Google Engineer Builds $1,500 Scanner to Rival $100,000 Robotic Rigs

Linear Book Scanner — Google TechTalks

Google TechTalks

Replacing robotic page-turning arms with a stationary vacuum prism and off-the-shelf contact sensors cuts automated book digitization costs to $1,500, making mass preservation viable despite minor handling defects.

AI-assisted summary of Linear Book Scanner — original source. Check the original for full context.

Source published:

Key ideas

  • Dany Qumsiyeh has engineered a low-cost, non-destructive automatic book scanner that shuttles a book over a stationary prism using vacuum slots and CIS sensors instead of robotic arms. (3:19)
  • The operational prototype costs roughly $1,500 to fabricate—about half of which came from an off-the-shelf donor duplex scanner—compared to commercial automated scanners costing upwards of $100,000. (18:48)
  • The machine scans a 300-page book in approximately 30 minutes at 11 seconds per two-page cycle, operating autonomously enough that a single worker could supervise ten machines simultaneously. (18:48)
  • Initial testing across 50 books revealed an operational success rate of around 60%, with 45% of tested volumes recording at least one skipped page and 45% showing at least one small fold or tear. (21:19)

Detailed notes

## Background Presented by Dany Qumsiyeh, an engineer on the Google Books team, this talk was delivered at a Google TechTalk on May 3, 2012, and published to YouTube by Google TechTalks on November 12, 2012. The session introduces a hardware prototype of an automatic, non-destructive book scanner developed in 20% time and constructed in the Google Workshops to digitize books at low cost without destroying their bindings.

## TL;DR Dany Qumsiyeh demonstrates the Linear Book Scanner, an innovative non-destructive prototype that turns and captures pages using a continuous shuttle mechanism, vacuum suction, and off-the-shelf contact image sensors (CIS). Costing around $1,500 compared to $100,000 commercial alternatives, the machine bypasses complex multi-axis robotics by sliding a book face-down across a stationary prism that separates pages using gentle air currents and vacuum channels. While currently constrained by motor torque, gutter dead zones, and occasional feed errors on damaged paper, the system automates digitizing a 300-page volume in roughly 30 minutes, allowing a single operator to manage multiple units simultaneously.

## Atmosphere & style The presentation carries the pragmatic, enthusiastic spirit of a classic engineering show-and-tell, balancing self-deprecating hacker ingenuity with rigorous mechanical problem-solving. Qumsiyeh presents with understated, modest clarity, drawing chuckles as he recounts grinding circuit boards on belt sanders and building the first iteration out of office dumpster cardboard and an air mattress pump. The exchange feels intimate and collaborative, as curious technical peers probe edge cases, sensor physics, and library preservation standards in a genuinely supportive room.

## Bulletin - [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) Dany Qumsiyeh has engineered a low-cost, non-destructive automatic book scanner that shuttles a book over a stationary prism using vacuum slots and CIS sensors instead of robotic arms. - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) The operational prototype costs roughly $1,500 to fabricate—about half of which came from an off-the-shelf donor duplex scanner—compared to commercial automated scanners costing upwards of $100,000. - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) The machine scans a 300-page book in approximately 30 minutes at 11 seconds per two-page cycle, operating autonomously enough that a single worker could supervise ten machines simultaneously. - [21:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1279s) Initial testing across 50 books revealed an operational success rate of around 60%, with 45% of tested volumes recording at least one skipped page and 45% showing at least one small fold or tear.

## Key takeaways - [00:24](https://www.youtube.com/watch?v=4JuoOaL11bw&t=24s) Existing non-destructive book scanners face a steep divide between labor-intensive manual overhead scanners and high-end robotic systems that cost around $100,000. - [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) The Linear Book Scanner replaces complex robotic arms with continuous back-and-forth linear motion of a book cradle across a stationary triangular prism. - [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) Vacuum suction from a standard household cleaner pulls the bottom-most leaf through a central channel while an air jet fluffs incoming pages to prevent multi-sheet feeds. - [06:00](https://www.youtube.com/watch?v=4JuoOaL11bw&t=360s) Straight diagonal channels force the paper to twist and rip against its spine binding; widening the channel vertically near the top allows the page to peel freely without destructive tension. - [12:54](https://www.youtube.com/watch?v=4JuoOaL11bw&t=774s) Using contact image sensors (CIS) rather than cameras provides full RGB data for every pixel at 300 DPI without Bayer filter interpolation, at a raw sensor cost of around $20. - [09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) To capture text close to the inner gutter binding, the square housings and circuit boards of off-the-shelf CIS sensors were modified on a belt sander down to a 9 mm margin. - [09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) Opening the prism geometry beyond 60 degrees was rejected because paper drapes poorly at wider angles and library conservation standards discourage flattening book spines. - [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) Automation relies on an MDrive stepper motor that streams an ASCII newline per stroke directly over serial into a piped open-source SANE scanimage command. - [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) An integrated infrared break-beam sensor monitors page transit through the vacuum channel, automatically halting the drive motor if a leaf fails to turn or arrives late. - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) Total hardware costs sit at approximately $1,500, though custom driver boards and flat sheet-metal fabrication could lower volume manufacturing costs even further. - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) Cycle speed is currently limited to 11 seconds per page pair by stepper motor torque constraints when handling heavier volumes. - [25:47](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1547s) The prototype cannot yet automatically detect double feeds where two leaves stick together, relying on post-scan UI mosaic inspection of page numbers to flag skipped sheets. - [27:30](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1650s) Preservation experts endorse low-cost automated scanning because the broader societal benefit of digital access outweighs the small risk of occasional paper damage.

## Detailed walkthrough ### [00:24](https://www.youtube.com/watch?v=4JuoOaL11bw&t=24s) Introduction and the Existing Landscape of Book Scanners

Jeff Breidenbach introduces Dany Qumsiyeh, an engineer on the Google Books team, to present a hardware prototype of an automatic book scanner. Breidenbach jokes about how the presentation came together through word of mouth: “This talk began with an idea of showing this device to a single person. And that person invited two more people. And you're all here through the incredible power of recursion.”

Dany powers down the machine on stage and frames the project by reviewing current book scanning methods:

```mermaid flowchart TD accTitle: Comparison of Book Scanning Approaches accDescr: Overview showing destructive, manual, commercial robotic, and low-cost DIY/linear approaches to book scanning.

subgraph Destructive["High Speed / Destructive"] D1["Cut off binding"] --> D2["Sheet-fed scanner"] end

subgraph ManualNonDestructive["Manual Non-Destructive"] M1["Manual page turn"] --> M2["Glass platen press"] M2 --> M3["Camera capture"] M4["DIY Book Scanner Community\n(diybookscanner.org)"] end

subgraph CommercialRobotic["Commercial Robotic"] R1["Robot suction arm\n(e.g., Kirtas)"] R2["V-shaped linear sensor\n(e.g., Treventus: ~$100k)"] end

subgraph LinearScanner["Linear Prototype"] L1["Prism shuttle + Vacuum + CIS scan bars"] end ```

The simplest method is destructive: slice off the binding and run the loose sheets through a high-speed sheet-fed scanner. While the scan quality is pristine, it ruins the book. Non-destructive options, however, have historically faced steep trade-offs between manual labor and prohibitive cost: - **Manual Overhead Scanners:** Operators turn each page by hand, lower a glass plate to press the pages flat, and trigger overhead cameras. While relatively fast, it demands constant human labor. A large enthusiast community at `diybookscanner.org` builds these manually using cardboard, wood, and off-the-shelf parts. - **Robotic Arm Scanners (e.g., Kirtas):** These automate the human movement by using a robotic suction arm that picks up individual pages and flips them over. - **V-Shaped Robotic Scanners (e.g., Treventus):** Treventus uses a mechanized V-shaped head that moves up and down into the book cradle. It applies vacuum suction to hold two pages against the platen while linear sensors along the tip capture images. Qumsiyeh highlights both their reputation and their primary drawback: “It's really cool. And these are state of the art machines. And they've gotten a lot of praise for being gentle to books. The main issue is that they're really expensive. One of these Treventus machines can cost on the order of $100,000.”

[03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) The Linear Book Scanner Prototype and Vacuum Separation

Dany shows a close-up video of his custom-built prototype in action. The operational principle relies on continuous linear motion over a stationary scanning head rather than moving cameras or complex robotic limbs. A book rests face-down atop a prism-shaped rig and shuttles back and forth. Inside the prism, linear sensors—identical to those found in standard flatbed or sheet-fed scanners—capture page images as the paper glides across them.

Air suction is provided by an ordinary household vacuum cleaner attached beneath the rig. As the book shuttles over the prism, vacuum slots pull the bottom-most leaf through a central channel to deposit it on the opposite side. An audience member asks how the machine avoids picking up multiple pages at once. Dany explains that as the stack of pages slides over the curved surface, suction is applied only to the single page directly in contact with the vacuum slots: “The other pages don't actually receive any suction, so they will tend to stay straight and go across.” To prevent sheets from sticking due to friction or static, the prototype incorporates an upward blast of compressed air targeted at the page edges, gently fluffing and separating them before they hit the dividing metal plate.

Because the prototype avoids complex multi-axis robotics, Dany emphasizes its affordability and elegance: “So it's really simple. And it doesn't involve very many parts, so it's not very expensive to build.”

[06:00](https://www.youtube.com/watch?v=4JuoOaL11bw&t=360s) Channel Geometry and Preventing Page Tearing

Dany delves into the subtle geometry of the internal channel, walking through earlier failed designs to show why a simple cut does not work:

```mermaid flowchart LR accTitle: Evolution of Channel Geometry accDescr: Three iterations of internal channel design showing progression from simple diagonal to relief cutout to vertical expansion.

subgraph Iteration1["1. Straight Diagonal Slot"] I1A["Page enters diagonal"] --> I1B["Binding stays rigid"] I1B --> I1C["Page forced to twist & tear"] end

subgraph Iteration2["2. Flat Intermediate Zone"] I2A["Enters diagonal slot"] --> I2B["Flat central relief"] I2B --> I2C["Page swings across untwisted"] end

subgraph Iteration3["3. Vertical Expansion"] I3A["Diagonal vacuum pickup"] --> I3B["Top opening expands vertically"] I3B --> I3C["Relieves binding stress without tearing"] end ```

Cutting a basic diagonal slot through the prism causes immediate failures: while the book's spine moves in a straight line, the middle of the sheet inside a simple diagonal slot is forced to twist along the diagonal plane, which paper cannot tolerate without ripping. To prevent this, the channel incorporates a flat section in the middle of the diagonal path, giving the leaf an unconstrained intermediate zone where it can swing freely across without twisting.

A second mechanical failure emerged in an early cardboard prototype involving the angle of the vacuum pull. When vacuum was pulled along a straight diagonal channel all the way to the top, the diagonal bend traveled directly into the book’s spine: - As the book moved laterally, the sheet was forced to bend sharply against its own binding. - To keep advancing, the paper had to shear itself directly off the active vacuum slots. - When the lateral resistance exceeded the tensile strength of the paper, the page tore.

To solve this, Dany modified the channel profile: the slot begins as a diagonal opening near the bottom to catch and curve the leading corner, but widens into a vertical opening toward the top. Once the page reaches this expanded vertical section, it loses lateral contact with the right-hand support surface and can swing freely through the void, peeling off the vacuum zone with minimal mechanical force on the binding.

[09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) Contact Image Sensors (CIS) and Scanning Geometry

Dany holds up the scanning bar used in the machine: a contact image sensor (CIS) pulled directly from an off-the-shelf sheet-fed scanner. The sensor contains an array of photodetectors and pulses RGB illumination to capture full color.

By salvaging the internal scan bars and driver boards from commercial scanners, Dany bypassed the need to engineer custom readout electronics or write custom drivers: “You can tell the computer to scan and it thinks the scanner is just a regular Canon scanner. And it captures images. So there's very little software required. And we didn't have to do any of the electronics.”

```mermaid flowchart TD accTitle: Contact Image Sensor Trade-offs accDescr: Diagram illustrating the physical packaging of CIS versus optical CCD sensors in relation to the book spine groove.

subgraph CISHack["CIS Sensor (Current Prototype)"] C1["Square casing cannot reach apex"] C1 --> C2["Modified on belt sander"] C2 --> C3["Achieves ~9 mm gutter margin"] end

subgraph CCDAlternative["Folded Optics CCD (Production Alternative)"] O1["Linear sensor with optical lens"] O1 --> O2["Focal path views directly into sharp apex corner"] O2 --> O3["Zero gutter loss"] end ```

However, the physical housing of off-the-shelf CIS bars creates a physical packaging constraint: - **Square Sensor Packaging:** The square cross-section of the CIS casing prevents the sensing element from seating all the way into the sharp 60-degree peak of the prism, creating a dead zone where text deep in the gutter (inner margin) cannot be imaged. - **The Belt Sander Modification:** To mitigate this, Dany ground down the edges of the sensor housings on a shop belt sander: “And it turns out you can grind away quite a bit of that circuit board before the thing stops working. And that was enough to get us within about nine millimeters of the groove, which is OK for most books.” - **Production Alternative (CCD Line Sensors):** For a production-grade machine, Dany suggests replacing CIS bars with line-scan CCDs using mirrors and focusing lenses, allowing the optical path to inspect the vertex directly without mechanical obstruction.

In response to an audience member asking whether widening the prism angle (such as opening it to 90 degrees) would allow the CIS bars to tuck closer to the spine, Dany explains why that introduces new problems: - Paper does not drape completely flat when forced open over wide angles, meaning page curl near the gutter would still lift away from the focal plane and obscure text. - Library preservation standards strictly prohibit opening rare or fragile bindings too wide: “Also, librarians generally like to keep the-- they consider it very bad to open the book too far. And so if you can keep the angle of opening as small as possible, that's better for the book.”

Dany concludes the segment by presenting a sample scan captured on the rig at a default resolution of 300 DPI, noting that the resulting image is flat, crisp, and identical in visual fidelity to an image produced by a standard desktop flatbed scanner.

[12:54](https://www.youtube.com/watch?v=4JuoOaL11bw&t=774s) Sensor Selection and Motion Mechanics

Dany Qumsiyeh explains the advantages of linear contact sensors over conventional digital cameras for document scanning. While standard cameras may boast high megapixel counts, each pixel relies on a Bayer filter capturing only red, green, or blue. In contrast, linear scanner heads capture full RGB color data for every single pixel at 300 DPI, yielding vastly superior image quality at a fraction of the price: “And these sensors are only like $20, compared to a very expensive camera that you would need.”

```mermaid flowchart LR accTitle: Linear Scanner Architecture accDescr: Diagram showing the physical movement, optical sensor calibration, and drive mechanisms of the linear book scanner.

subgraph Motion ["Drive System"] M["MDrive Stepper Motor"] -->|"drives"| TB["Timing Belt"] TB -->|"actuates"| S["Saddle Cradle"] end

subgraph Sensing ["Scanning & Calibration"] S -->|"passes over"| CIS["Linear CIS Sensors"] WS["White Calibration Strips"] -.->|"covers at home"| CIS end end ```

Addressing an audience question regarding motion linearity and whether speed variations distort the scans, Qumsiyeh reports that image output shows no noticeable disturbances. The carriage motion relies on a stepper motor coupled to a standard scanner timing belt, ensuring highly repeatable travel. Although early design phases explored optical pattern tracking on the saddle to correct for speed irregularities, the native mechanical precision proved sufficient.

The moving carriage, referred to as the "saddle," incorporates adjustable clamps to support the book. These clamps serve a vital structural purpose: as scanning progresses toward the start or end of a thick volume, the asymmetrical weight distribution causes unsupported pages to sag and bow near the gutter binding, which impedes page separation or causes the volume to tilt off balance. When asked whether softcover paperbacks can be scanned, Qumsiyeh confirms they work without issue, though the clamps naturally provide firmer registration against rigid hardcover boards. During setup, an operator places the book centrally on the saddle and adjusts the clamps until they make light contact with the covers.

- Scanner sensors capture true RGB color per pixel at 300 DPI, avoiding camera Bayer interpolation. - Linear sensor modules cost approximately $20 each. - White strips fixed to the home position of the saddle allow the sensors to calibrate white-balance values automatically before every scan run.

---

[15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) Automation Hardware and Control Software

The scanning carriage is propelled by an integrated MDrive stepper motor equipped with an on-board programmable microcontroller. The control firmware is exceptionally compact: users preprogram travel speed, endpoints, and reversal limits into the motor memory. Once powered on, the motor executes continuous back-and-forth passes autonomously without real-time external direction.

```mermaid flowchart TD accTitle: Software Automation Pipeline accDescr: Pipeline illustrating how newline characters from the motor trigger SANE scanimage capture and feed the viewer interface.

M["MDrive Motor (Serial Output)"] -->|"outputs newline on cycle start"| P["Unix Pipe / Shell"] P -->|"supplies Enter key"| SANE["SANE scanimage (--batch-prompt)"] SANE -->|"saves scanned page"| V["Viewer Software (Mosaic UI)"] BB["Break-Beam Sensor"] -->|"halts on misfeed"| M ```

To coordinate page capture with carriage travel, the MDrive outputs a single ASCII newline character over an RS-232 serial interface each time it reaches the start of a stroke. The host machine runs `scanimage`, a command-line tool from the open-source SANE (Scanner Access Now Easy) project, configured in batch-prompt mode (`--batch-prompt`), which waits for an Enter keystroke to trigger acquisition: “And so then all you have to do is take the line breaks from the motor and pipe them over to scanimage.” This entire coordination pipeline collapses into a single piped shell command.

For quality assurance, software written by Google colleague Jeff provides real-time visualization of incoming page pairs. The interface features an inspection view that extracts and aligns the page numbers of each scanned page into a vertical mosaic: - Operators verify scanning integrity at a glance by scanning down sorted numerical columns (e.g., 361, 341, 321, 301) to spot skipped or duplicated leaves instantly. - A physical infrared break-beam optical sensor is positioned directly inside the page-turning vacuum throat. - The break-beam sensor wires straight into the MDrive microcontroller input lines; if a page arrives early, late, or fails to appear, the motor immediately halts carriage travel to avert damage.

---

[18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) Prototype Economics, Scan Rates, and Motor Constraints

The demonstrated device represents the project's third iteration and its very first rigid metal build. Qumsiyeh recounts that the original proof of concept was built out of scrap cardboard retrieved from an office dumpster and powered by an inflatable air mattress pump, which surprisingly proved capable of separating and turning pages.

```mermaid flowchart TD accTitle: Prototype Bill of Materials Breakdown accDescr: Cost distribution of the prototype hardware showing a total cost of 1,500 dollars.

Total["Total Prototype: ~$1,500"] Total --> CIS["Torn-down Flatbed Scanner: ~$750"] Total --> Mech["Vacuum Pump, Motor & Machined Metal: ~$750"] ```

The current metal unit takes roughly 40 seconds to clamp and index a new volume. The cycle time is 11 seconds per page pair, which equates to approximately 30 minutes for a standard 300-page book. While slower than industrial book digitization rigs, Qumsiyeh emphasizes the economic leverage of complete automation: “This is a lot slower than many of the alternatives. But since it's fully automatic, one operator should be able to run like 10 of them.” Total hardware costs run around $1,500, with half of that figure accounted for by the donor desktop flatbed scanner stripped down for its optical sensors and electronics.

In a brief Q&A exchange regarding speed and safety, Qumsiyeh clarifies operational constraints: - Higher-end variants of the MDrive stepper motor include closed-loop stall detection and torque sensing, which could halt the cycle if paper snagging increases mechanical resistance. - Cycle throughput is currently constrained by motor torque: running heavier volumes at higher velocities stalls the existing motor. - Speed can be improved either by installing a more powerful motor or by compacting the physical geometry to shorten total carriage stroke travel.

---

[21:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1279s) Paper Handling Limitations, Error Rates, and Future Directions

Testing across a test library of roughly 50 books revealed several mechanical limits. Approximately 60% of arbitrary off-the-shelf books can physically run through the rig. Books must fit within the saddle clearance and have pages long enough to engage the vacuum suction slots. Paper weight also introduces challenges: very thin leaves (such as India paper in dictionaries) pull too easily into the vacuum slots, while heavy cardstock covers lack the flexibility to negotiate the turning bend. Regulating vacuum suction pressure is essential for tailoring the machine to diverse paper weights.

Error rates on the initial 50-book evaluation highlight the need for further refinement: - Around 45% of test volumes suffered one or two skipped pages across the entire book. - Around 45% experienced a fold or small tear on at least one page. - Qumsiyeh clarifies to the audience that these percentages represent the frequency of volumes that suffered at least one defect, rather than aggregate per-page failure rates: “Of the books. Yes. Not the pages. So it's only a couple errors in the book.” - Most page tears occurred on leaves with pre-existing edge nicking that caught in the slot; modern runs using the break-beam sensor catch these jams and stop the motor before folds worsen.

Future iterations will explore alternative vacuum orifice geometries, replacing bare aluminum components with stainless steel to prevent aluminum oxide rub-off along page edges, and incorporating CCD sensors capable of scanning tight into the inner gutter binding. For high-volume facility scaling, Qumsiyeh proposes chaining multiple V-shaped linear scanners in a continuous diagonal line, allowing two operators stationed at opposite ends of a conveyor to scan a continuous stream of books.

[25:03](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1503s) Project Acknowledgments and Opening the Floor

Dany Qumsiyeh concludes the presentation segment by recognizing key collaborators who brought the Linear Book Scanner prototype to fruition. He credits a crucial collaborator who authored the entire software stack, helped source and configure the sensors, installed the break-beam sensor, and assisted in dismantling commercial scanners for parts. Qumsiyeh notes that beyond technical labor, this teammate provided essential advice and moral support throughout the development process. He also expresses gratitude to his current engineering manager, Stephane, and his former manager, Ram, for greenlighting the initiative as a 20% project and offering continual support. With the acknowledgments wrapped up, Qumsiyeh invites the audience to ask questions.

[25:47](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1547s) Error Handling: Page Jams, Double Feeds, and Unit Economics

The first audience question addresses error recovery: what occurs when something goes wrong during a scan. Qumsiyeh explains that the current prototype detects when a page fails to turn properly; once triggered, “the machine just stops,” allowing the operator to inspect the issue, reset the mechanism, and restart the cycle. However, the prototype currently lacks a mechanism to detect double feeds where two pages stick together and travel through simultaneously, causing the intermediate page to be skipped. Qumsiyeh suggests that adopting techniques from commercial sheet-fed document scanners could solve this problem in future revisions.

```mermaid flowchart TD accTitle: Linear Book Scanner Error Handling Flow accDescr: Flowchart showing how the prototype handles single-page stalls versus undetected multi-page feeds.

subgraph Monitoring ["Sensory Detection"] A["Page transport across head"] --> B{"Break-beam sensor status"} end

subgraph Resolution ["Failure Modes & Handling"] B -->|"Page fails to turn"| C["Automated system halt"] C --> D["Operator clears jam & resets run"] B -->|"Two pages stick together"| E["Double feed passes undetected"] E --> F["Missed capture of inner page"] F -.->|"Future revision"| G["Integrate sheet-fed multi-page sensors"] end ```

When asked whether the system could be manufactured even more cheaply—given that the scanner sub-assembly represents the largest expense—Qumsiyeh affirms that costs could drop significantly with custom electronics. The raw optical sensor components used in the build cost approximately $20 each. Fabricating custom driver boards at volume would lower hardware expenses well below off-the-shelf scanner tear-downs. Under those conditions, the dominant manufacturing expense would simply be assembly labor. He emphasizes that the structural fabrication remains inherently inexpensive because the chassis components are flat sheet metal profiles that are trivial to cut.

- When a page fails to clear the break beam, the scanner automatically halts to prevent further mechanical damage. - Multi-page feeds remain an unaddressed limitation in the current revision, requiring future integration of thickness or ultrasonic double-feed sensors. - Raw linear image sensor components cost roughly $20 per unit. - The mechanical chassis consists primarily of 2D flat-cut metal pieces, keeping fabrication overhead low.

[27:30](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1650s) Dirt Streaks, Page Safety, and Library Risk Acceptance

An audience member asks about the risk of dirty volumes leaving smudges or continuous lines down the optical sensors, followed by a query regarding where the machine was manufactured. Qumsiyeh shares that the prototype was built in the on-campus Google Workshops. Addressing debris, he admits that smeared grime on a page could adhere to the glass sensor bar and streak successive scans, necessitating a detection and cleaning routine. However, he jokes about an accidental advantage of the suction system: “we joke that since the book is travelling over these vacuum slots, we're actually cleaning out dirt from the book as it goes. And that's a feature.”

The discussion then turns to paper preservation and whether the linear design risks tearing pages compared to commercial robotic scanners. Qumsiyeh counters that even high-end, dedicated page-turning machines occasionally tear paper, though it remains rare enough for institutional use. He recounts consulting with a preservation specialist from UC Berkeley, who strongly supported the concept:

> “And he was really excited basically that the benefit of having digital versions of these for people to use was greater than the small risk of causing damage to some books.”

Qumsiyeh explains that libraries already accept inherent wear and tear as a standard operational trade-off: when patrons check out physical books, the items inevitably return in degraded condition, yet institutions accept that reality because enabling public access to the content justifies the minor wear. His long-term goal is to refine the linear mechanism until it is gentle enough for rare and fragile collections.

[29:16](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1756s) Optical Geometry, Hardware Architecture, and Project Future

In response to a question asking why the scanning sensors sit at a distance from the page-turning vacuum slot rather than directly adjacent to it, Qumsiyeh explains the mechanical layout. The team chose to keep both scan bars aligned in a single physical assembly so both page faces are scanned simultaneously. The exit vacuum slot actually extends upward to meet the scanner assembly. Separating the scan bars onto opposite sides of the intake and exit slots would permit placing each sensor closer to the turn point, thereby reducing the travel distance required for each leaf.

Clarifying the optical hardware source, Qumsiyeh explains that the setup is not two independent retail scanners, but rather the harvested components of a single duplex sheet-fed document scanner. The original donor machine used one unified circuit board wired to two contact image sensor (CIS) bars—one for the front and one for the back of the fed sheet. Finally, when asked about the future roadmap and next steps for the Linear Book Scanner project, Qumsiyeh demurs on making public forecasts, noting, “I don't know. But we'll talk a little bit more about that later.” At this point, host Jeff Breidenbach steps in to wrap up the recorded session, inviting the audience to applaud Dany Qumsiyeh.

## Mentioned Dany Qumsiyeh, Jeff Breidenbach, Google Books, Google Workshops, diybookscanner.org, Kirtas, Treventus, Canon, MDrive, SANE (Scanner Access Now Easy), scanimage, UC Berkeley

## Video description Google Tech Talk May 3, 2012

Presented by Dany Qumsiyeh

ABSTRACT

See a hardware prototype of an automatic, non-destructive book scanner. The machine turns pages automatically, and captures high-resolution images of each page. It was developed in 20% time, and built in the Google Workshops.

There will be a live demo and a short technical presentation, describing the design process.

Speaker Info: Dany Qumsiyeh is an engineer on the Google Books team.

For more information on this project, see: http://code.google.com/p/linear-book-scanner/

[Source](https://www.youtube.com/watch?v=4JuoOaL11bw)

中文简报

谷歌工程师打造1500美元扫描仪，媲美10万美元级机器人设备 用固定真空棱镜和现成接触式传感器取代机器人翻页臂，将自动化图书数字化成本降至1500美元，尽管存在轻微处理缺陷，但使大规模图书保护变得可行。 达尼·库姆西耶展示了线性图书扫描仪（Linear Book Scanner），这是一种创新的无损原型机，利用连续穿梭机构、真空吸附和现成接触式图像传感器（CIS）来实现翻页与采集。相比成本达10万美元的商业替代方案，该设备造价仅约1500美元；它避开了复杂的多轴机器人系统，通过让图书朝下在固定棱镜上滑动，并利用温和气流和真空通道分离书页。尽管目前受到电机扭矩、装订死区以及损坏纸张上偶发进纸错误的限制，该系统仍可在约30分钟内完成一本300页图书的自动化数字化，从而允许一名操作员同时管理多台设备。 达尼·库姆西耶研发了一款低成本、无损的自动图书扫描仪，利用真空吸槽和CIS传感器代替机械臂，使图书在固定的棱镜上方穿梭移动。 (3:19) 这台可运行的原型机制造费用约为1500美元——其中大约一半成本来自一台现成的供体双面扫描仪——而商业自动化扫描仪的造价则高达10万美元以上。 (18:48) 该设备以每双页11秒的速度运行，扫描一本300页的图书约需30分钟，其自主运行程度足以让一名工人同时监管10台机器。 (18:48) 在对50本书进行的初步测试中，运行成功率约为60%，其中45%的受测书卷记录到至少出现一次漏页，45%出现至少一次轻微折叠或撕裂。 (21:19)

Reading summary

Dany Qumsiyeh, an engineer on the Google Books team, presents an automated linear scanner prototype that replaces costly robotic digitizers with simple vacuum mechanics and low-cost components. By translating an open book across a stationary prism equipped with suction slots, compressed air, and salvaged scan bars, the design automates non-destructive scanning at a fraction of commercial expense. Vacuum-Assisted Page Separation and Channel Geometry (3:19) The prototype dispenses with multi-axis robotic arms in favor of continuous linear motion across a stationary prism. As an inverted book shuttles back and forth on a moving saddle, vacuum slots pull the bottom leaf through a central channel to turn it. To avoid pulling multiple leaves simultaneously, an upward blast of compressed air fluffs page edges before they reach the channel, while vacuum suction acts exclusively on the page in direct contact with the aperture. Early channel designs failed because cutting a simple diagonal slot forced the sheet to twist against its rigid binding, tearing the paper. Qumsiyeh addressed this by adding a flat relief zone in the middle of the channel, followed by an opening that expands vertically toward the top. This geometry allows the leaf to swing freely through the void, peeling off the vacuum slots with minimal mechanical stress on the spine. Contact Image Sensors and Optical Packaging (9:43) Rather than employing overhead cameras with Bayer-filter interpolation, the scanner uses Contact Image Sensors (CIS) salvaged from a duplex sheet-fed scanner. These capture true RGB color at 300 DPI for roughly $20 per sensor. Repurposing the donor scanner's original driver boards allows standard operating software to drive image acquisition through off-the-shelf drivers. Carriage movement relies on an MDrive stepper motor and timing belt, while white strips at the saddle home position handle automatic calibration before each scan. However, commercial CIS casings present physical packaging challenges. Their square cross-sections prevent sensor elements from seating into the sharp apex of the 60-degree prism, obscuring gutter text. Qumsiyeh ground down the circuit boards on a belt sander to image within 9 mm of the spine. Widening the prism angle was ruled out because paper curls away from the focal plane over wide angles, and preservation standards prohibit opening delicate bindings too far. For production machines, Qumsiyeh recommends line-scan CCDs with folded optics to view the vertex without obstruction. Software Coordination and Break-Beam Safety (15:55) Automation operates through a streamlined software and firmware loop. An MDrive stepper motor is preprogrammed with cycle endpoints and speeds, running autonomously without real-time host control. At the start of each stroke, the motor outputs an ASCII newline over RS-232. A Unix shell pipes this character into the open-source SANE utility scanimage running in batch-prompt mode, triggering capture with a single command line. A companion viewer extracts scanned page numbers into an aligned vertical mosaic, allowing operators to verify sequencing at a glance. To protect books, an infrared break-beam sensor inside the vacuum throat connects directly to the motor microcontroller. If a page fails to pass through the throat on schedule, the motor halts immediately, preventing folds or tears before the cycle can worsen the jam. Prototype Economics, Throughput, and Error Rates (18:48) The prototype metal unit cost approximately $1,500—half of which covered the donor scanner—compared to commercial robotic systems like Treventus that cost roughly $100,000. Qumsiyeh projects that custom driver boards and flat-cut sheet metal chassis would lower production costs further. Throughput sits at 11 seconds per two-page spread, taking roughly 30 minutes for a 300-page book. While slower than industrial machines, Qumsiyeh emphasizes that full automation allows one operator to oversee up to ten scanners simultaneously. Testing across 50 books revealed that approximately 60% of arbitrary volumes could be handled. However, 45% of test volumes suffered one or two skipped pages across the book, and 45% experienced a fold or small tear, typically on pages with pre-existing edge damage. Regarding preservation, Qumsiyeh cites a UC Berkeley preservation specialist who agreed that the public access enabled by digitization outweighs the low risk of occasional wear, mirroring the accepted risks of library patron lending. Key mechanical limitations remain to be resolved. The prototype cannot detect multi-page double feeds, struggles with very thin or rigid paper weights, and leaves a 9 mm dead zone near the gutter. Future development goals include integrating sheet-fed double-feed sensors, replacing aluminum surfaces with stainless steel to prevent page marking, and adopting folded-optics CCDs.

精炼阅读摘要

Google 图书团队工程师 Dany Qumsiyeh 研发了一款基于线性往复运动与真空吸力翻页的非破坏性图书扫描仪原型，旨在大幅降低传统自动化扫描设备的昂贵制造成本。 现有图书扫描技术的成本与效率瓶颈 (0:24) 传统图书扫描方式长期受制于效率、破坏性与设备成本的多重权衡。最具成本效益的方案是切除书脊的破坏性进纸扫描，但这会损毁书籍本身；而在非破坏性领域，人工扫描仪需要操作人员手动翻页并按压玻璃板，依赖极高的人力投入。以 Treventus 为代表的高端自动化 V 型扫描仪虽兼具温和度与自动化，但售价高达约 10 万美元，使普通机构难以承受。 线性图书扫描仪（Linear Book Scanner）原型摒弃了复杂的多轴机械臂设计，改用书籍沿固定棱镜式夹具往复滑动的线性机制。内部通过家用真空吸尘器提供负压，配合边缘压缩空气吹拂来分离纸张，确保仅有最底层的单页被吸入通道翻转，实现了大幅简化机械结构的低成本非破坏性扫描路径。 通道内部几何结构演进与防撕裂设计 (6:00) 设计初期的简单对角线切槽导致了严重的技术问题。由于书脊保持直线行进，纸张在斜切通道内会被迫产生扭曲，极易扯破。为解决这一应力集中现象，设计团队在对角线路径中部加入了平整过渡区，为纸张提供自由摆动的间隙。 此外，早期的全斜角真空吸附会使纸张在行进时强行沿书脊自身折弯并与吸附槽产生横向剪切，撕裂风险极高。团队随即将通道剖面改进为“底部对角吸附引偏、顶部垂直扩宽”的构型。纸张进入垂直扩展段后即可脱离侧向支撑面并顺利脱离真空槽，将书脊处的机械牵拉应力降至最低。 接触式图像传感器（CIS）改造与扫描光学特性 (9:43) 原型直接拆解了一台市售双面进纸扫描仪，利用其接触式图像传感器（CIS）及控制主板。这种传感器通过脉冲 RGB 光源在 300 DPI 分辨率下捕捉每个像素的完整三原色数据，避开了普通相机拜耳滤镜的插值损失，且单根传感器元件成本仅约 20 美元。更重要的是，计算机可直接通过常规驱动将其识别为标准扫描仪，大幅减少了自定义电子控制与驱动编写的负担。 然而，商用 CIS 传感器的方形封装无法紧密贴合棱镜尖端 60 度的夹角，会导致书缝（内侧边距）区域出现成像盲区。Qumsiyeh 使用砂带机打磨了传感器外壳与部分电路板边缘，使盲区缩窄至约 9 毫米。针对生产级设备，他建议采用结合反射镜的线阵 CCD 光学方案，彻底消除书缝遮挡；同时因受限于图书馆对古籍最大张开角度的严格限制，无法通过将棱镜展平至 90 度来简单规避该问题。 步进驱动、自动化流程控制与质量检测 机身运动由一台集成控制芯片的 MDrive 步进电机配合标准同步带驱动，内置固件独立控制行进速度与往复边界。电机每次到达行程起点时，会通过 RS-232 串口向主机发送一个换行符，以此触发开源软件 SANE 的命令行工具 scanimage 执行批处理捕捉，仅凭一条 Unix 管道命令便完成了硬件运动与软件图像采集的协同。 系统在翻页喉管内布置了红外穿透光束传感器，与电机控制引脚直连；一旦纸张卡滞、提前或延迟到达，电机会立即急停以防书籍受损。配套软件还提供可视化拼接界面，能将各页码区域垂直排列成马赛克阵列，便于操作员一目了然地快速核对是否存在跳页或重复翻页现象。 设备经济性、故障表现与适用性限制 (21:19) 该金属原型的整机硬件成本约为 1,500 美元，其中约一半来自拆解的双面扫描仪元件；若未来采用定制电路板并在大批量下采购单价约 20 美元的裸传感器，成本仍将大幅下探。当前系统双页扫描耗时约 11 秒，处理一本 300 页图书需约 30 分钟。由于实现了全自动运行，一名操作人员理论上可同时监控并照看多台机器。 在针对约 50 本日常书籍的实测中，该设备的图书通过率约为 60%。非常薄的纸张容易过度凹陷，过厚的卡纸封面则缺乏弯曲韧性；在全部测试书籍中，约 45% 的书本出现过 1 至 2 页跳页，约 45% 出现过单页折痕或微小撕裂（多为带有原有裂口的纸页）。此外，当前原型尚缺乏多页粘连进纸的检测能力，但 Qumsiyeh 引用伯克利分校古籍修复专家的观点指出，相较于让文献数字化造福大众的收益，承担轻微磨损风险在图书馆界通常是可以接受的折衷方案。 该线性图书扫描仪验证了一种极具性价比的非破坏性数字化路径，但受限于进纸容错率、重张检测能力以及对超薄或硬卡纸的适应力。未来仍需进一步测试优化真空孔径几何形态、引入不锈钢材质以防止铝粉污染，并评估 CCD 光学元器件对深书缝采样的实用性。

Detailed notes — zh

## Background 该内容由 Google TechTalks 于 2012 年 11 月 12 日发布在 YouTube 平台上。视频记录了 Google Books 团队工程师 Dany Qumsiyeh 在 2012 年 5 月 3 日进行的技术演讲与现场演示，介绍了他利用 Google 的 20% 自由时间并在 Google 内部工坊制作的一台非破坏性自动线性图书扫描仪（Linear Book Scanner）硬件原型机。 ## TL;DR Dany Qumsiyeh 详细展示并讲解了他主导研发的低成本、非破坏性自动图书扫描仪原型机。该设备利用家用吸尘器产生的真空负压与精密设计的几何通道实现单页轻柔分离与自动翻页，并结合改装自消费级扫描仪的接触式图像传感器（CIS）实现双面高分辨率成像。演讲完整复盘了从纸箱胶带初号机到全金属第三代原型的迭代历程，客观分析了约 1,500 美元 BOM 成本下的吞吐能力，并坦诚探讨了漏页率、装订死区、纸张破损等局限性及图书数字化保护理念。 ## Atmosphere & style 整场演讲洋溢着浓厚硬核的 Google 极客创客文化氛围，语调务实、幽默且极具工程洞察力。主讲人 Dany 毫不避讳早期的挫折与原型机目前高达 45% 的样本瑕疵率，在被问及粗暴磨削电路板、把吸尘器当翻页引擎等细节时现场笑声不断。台下工程师观众的提问尖锐且专业，演讲在真诚坦率的技术切磋与对图书普及使命的共鸣中推向高潮，让人真实感受到车库式硬件创新那种粗粝却极具生命力的魅力。 ## Bulletin - [00:24](https://www.youtube.com/watch?v=4JuoOaL11bw&t=24s) Google 工程师 Dany Qumsiyeh 利用 20% 自由时间在 Google 工坊开发出全自动、非破坏性线性图书扫描仪原型机，BOM 原材料成本仅约 1,500 美元，远低于商用设备数万至十万美元的造价。 - [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) 该设备采用家用吸尘器负压结合特殊几何通道实现物理单页分离与无撕裂翻转，并复用拆解自普通馈纸式扫描仪的接触式图像传感器（CIS）完成 300 DPI 全彩扫描。 - [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) 软件端通过 Unix 管道将 MDrive 步进电机串口输出的换行符重定向至 SANE scanimage 命令，单行终端指令即可完成全自动循环采集与软硬件时钟同步。 - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) 当前样机扫描周期为每对页面 11 秒（300 页书约需 30 分钟），支持单人看管 10 台并发，但面临 45% 的书籍出现漏页或纸张轻微破损的可靠性瓶颈，仍需加装双页检测与材质改良。 ## Key takeaways - [00:24](https://www.youtube.com/watch?v=4JuoOaL11bw&t=24s) 现有图书数字化技术面临两难：切断书脊的破坏性馈纸扫描速度快但毁书，而 Treventus、Kirtas 等温和的非破坏性商用自动化扫描仪造价高达 10 万美元级别。 - [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) 扫描仪核心机械结构为棱柱形基座，书籍置于鞍座往复平移，利用普通家用吸尘器负压与底部吹气槽实现纸叠边缘松动与单页分离。 - [06:00](https://www.youtube.com/watch?v=4JuoOaL11bw&t=360s) 翻页通道从早期的对角线直线狭缝演进为复合几何开口（下部斜向吸入、上部垂直敞开），消除了纸张三维扭转应力并避免装订线在真空拉拽下撕裂。 - [09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) 通过在砂带机上手工打磨 CIS 传感器电路板外壳，团队将书脊装订槽的不可扫描死区极限压缩至约 9 毫米。 - [09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) 棱柱顶角若增大至 90 度虽便于布设传感器，但会导致纸页翘曲脱平且违背图书馆古籍保护标准，未来量产机更适宜采用配备缩放透镜的线阵 CCD 投影成像。 - [12:54](https://www.youtube.com/watch?v=4JuoOaL11bw&t=774s) 单价约 20 美元的 CIS 传感器每个物理像素均能真实捕获 RGB 三原色通道，在色彩扎实度与成本控制上显著优于数码相机的拜耳阵列插值方案。 - [12:54](https://www.youtube.com/watch?v=4JuoOaL11bw&t=774s) 鞍座两侧夹具不仅平衡了厚重书籍单侧堆叠时的重心倾斜，其初始复位位置上的白色校准条还能在扫描启动前自动校准基准白。 - [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) 上位机软件极度精简，仅凭管道重定向将 MDrive 电机串口换行符转给 SANE scanimage 命令即可自动运转，辅以页码马赛克视图快速抽检漏页。 - [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) 翻页吸槽内加装了对射式光电传感器（break beam sensor），能在翻页缺失或卡纸时向电机固件实时发送中断停机信号以保护纸张。 - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) 当前金属样机测试了约 50 本书，仅能稳定处理 60% 的书籍类型，超小开本无法覆盖吸槽，且约 45% 的书籍在某一页留有撕裂或折痕。 - [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) 铝合金导板摩擦会在纸缘留下灰色氧化铝微痕需升级为不锈钢材质，未来规划将单机串联构建如同流水线般的极高吞吐量扫描长廊。 - [25:41](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1541s) 当前原型无法检测两页粘连同时翻过的情况，未来需要引入商用馈纸式扫描仪常用的超声波或厚度传感器方案。 - [26:49](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1609s) 整机机械结构主要由扁平薄金属钣金激光切割而成，未来若自主设计传感器驱动电路，规模化量产最大的成本变量仅为组装人工工时。 - [28:12](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1692s) 伯克利图书保护专家对该设计持积极态度，认为公众获取全书数字化的巨大知识收益完全值得图书馆承担小概率的纸张磨损风险。 - [29:16](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1756s) 扫描模块实质拆解自一台市售商用双面馈纸扫描仪，利用其原装主板与双面对置的扫描棒实现单次通过同步采集书页正面与反面。 ## Detailed walkthrough ### [00:24](https://www.youtube.com/watch?v=4JuoOaL11bw&t=24s) 介绍与现有图书扫描技术概览 Google 工程师 Jeff Breidenbach 为本次演讲作了开场介绍。他风趣地提到，这场分享最初只是想向一个人展示这台设备，结果那个人又邀请了两个人，最终在递归的奇妙力量下，现场坐满了听众。随后，主讲人、来自 Google Books 团队的 Dany Qumsiyeh 接过话筒，在关闭运行中的机器后，开始介绍他利用 20% 自由时间在 Google 工坊中打造的自动图书扫描仪原型机。在深入讲解自己的设计之前，Dany 首先梳理了市面上现有的图书扫描技术，以此作为对比基准。 最简单的扫描方法莫过于切掉书脊装订线，然后将散页直接塞进馈纸式扫描仪中，这种方式图像质量极高，但弊端是彻底毁坏了书籍。因此，非破坏性图书扫描技术显得尤为关键。在非破坏性扫描领域，有一大类属于半自动设备：需要人工逐页翻书，通常会用玻璃压板将书页压平，再由顶部的相机抓拍双面图像。这种设备扫描速度尚可，但依赖大量的繁重体力劳动；在 diybookscanner.org 社区中，有许多爱好者利用纸板、木材等常见材料自行搭建此类设备并共享图纸。至于商用自动化扫描仪，像 Kirtas 这样的厂商采用机械臂模拟人手，通过气吸将单页翻动至另一侧；而 Treventus 公司则采用一种非常独特的 V 型升降结构，在上下移动时将书页吸附在两侧，并通过顶端集成的线性传感器直接扫出图像。 - 破坏性扫描：切断书脊后使用自动馈纸式扫描仪（Sheet-fed），速度快、质量高但不可逆毁坏书籍。 - 手动/半自动扫描：操作员抬起玻璃板、翻页、压平、相机拍摄，人工成本高昂（如开源社区 diybookscanner.org 的自制方案）。 - 商用机械臂扫描：例如 Kirtas，通过机械臂真空吸附页面进行翻页。 - V 型自动化扫描仪：例如 Treventus，采用真空吸附配合尖端线性传感器扫描，对古籍非常温和。 - Dany Qumsiyeh 在评价现有高端商业设备时指出：“The main issue is that they're really expensive.”（最主要的问题在于它们极其昂贵，一台 Treventus 机器的造价可高达 10 万美元级别）。 --- ### [03:19](https://www.youtube.com/watch?v=4JuoOaL11bw&t=199s) 线性扫描仪原型机架构与单页分离机制 Dany 展示了一段原型机运行的近景视频。这台机器的核心结构是一个棱柱形的基座（Prism），书籍平放在顶端来回往复平移。棱柱表面嵌有类似于普通平板扫描仪所使用的线性传感器，当书本滑过时实时捕获页面图像。在翻页环节，机器直接连接了一台普通的家用吸尘器提供负压抽气，将单张书页吸入特定通道并引导至另一侧。整个系统机械部件极少，因而整体造价非常低廉。现场观众随即提问：在高速移动过程中，如何确保每次只吸起并翻过一张纸？ ```mermaid flowchart TD accTitle: 单页吸附与翻页分离流程 accDescr: 展示书页堆沿曲面滑动、真空吸附底层单页、吹气分离以及金属挡片物理分流的过程。 subgraph Input["页面输入阶段"] A["书页堆沿曲面平移输入"] B["底部向上吹气松动纸页"] end subgraph Separation["真空吸附与物理分离"] C{"弯曲表面真空吸风槽"} D["底层单页紧贴曲面弯曲"] E["上层其余书页保持直线前行"] F["金属分离片物理隔开"] end subgraph Imaging["扫描与输出"] G["单页进入通道翻向对侧"] H["线性接触式传感器 CIS 逐行扫描"] I["电脑实时接收扫描图像"] end A --> B B --> C C --> D C --> E D --> F E --> F F --> G G --> H H --> I ``` 针对翻页防重叠的疑问，Dany 借助侧面示意图解释了单页分离的物理机理。成叠的书页从右侧沿曲面滑动，曲面特定位置开有真空吸气槽。由于负压直接作用于最底层的单张纸，只有这一张纸会被向下拉拽并贴合曲面弯折；上层的其余页面由于没有受到吸力，会依靠自身挺度保持平直向前滑动，随后两侧由一片金属分流片将其彻底隔开。考虑到纸张之间可能存在静电或受潮黏连，原型机还在纸叠边缘下方设置了向上吹拂的气流，利用气吹预先将纸页边缘吹散松动，有效避免了一次带入多页的情况。 - 运动方式：书籍放置在棱柱状结构上方，做水平往复直线运动。 - 真空动力源：直接使用常规家用吸尘器产生负压气流。 - 图像获取：利用内置的线性扫描模组实时采集页面。 - 翻页防重叠双重保障：底部垂直吹气松动纸张边缘，配合局部吸风槽仅吸附接触曲面的最底层单张纸。 - 面对观众关于如何精准分页的提问，Dany Qumsiyeh 直截了当地回应：“The question was how do you ensure that you only get one page? That's the next thing I'll talk about once this is done.” --- ### [06:00](https://www.youtube.com/watch?v=4JuoOaL11bw&t=360s) 翻页通道的几何拓扑演进与防撕裂设计 为了让观众理解翻页通道为何呈现出不规则的奇怪外观，Dany 详细复盘了设计过程中走过的弯路以及背后的几何力学原因。最直观的设想是在棱柱上直接切出一条贯通左右的对角线斜槽，让纸页顺着滑到对侧。然而测试表明，书脊必须保持水平直线运动，而处于斜槽中间的纸页却被迫发生空间扭曲（Twist）；纸张本身无法承受这种剪切扭力，极易直接撕裂。为此，改良方案在斜槽中间引入了一个平坦的过渡缓冲区（Intermediate flat section），使纸页滑入该区域后拥有自由摆动到对侧的空间，消除了扭转应力。 接下来遇到的第二个难题是真空脱附与书脊受力问题。在最初的硬纸板原型中，通道前表面仅开有一条笔直的对角线狭缝，真空吸力沿着斜线施加在纸张边角上。随着书本继续向左推进，沿对角线弯曲的变形区会一路延伸至装订书脊处；为了继续前进，纸张必须强行脱离真空吸附区域，但这会迫使纸张在书脊根部硬生生发生弯折。一旦脱离阻力过大，书页就会在书脊处被当场撕断。 最终成功的通道外形融合了复合几何结构：通道底部保持斜向开口以捕获页角，而在靠近顶部的区域则转变为向上垂直敞开的宽阔开口。当纸张被吸附并沿对角线弯曲推进到一定位置时，恰好进入顶部垂直敞开的无支撑区域。此时右侧已无实体表面阻挡，整张纸能够自由向对侧摆动，顺畅滑离真空吸附区，而不会给书脊脆弱的装订处施加破坏性的拉扯力。 - 直线对角槽缺陷：纸张在斜槽内发生三维扭转，导致纸张损坏。 - 中间平坦段改进：为纸页摆向对侧提供自由行程空间，消除扭矩。 - 直线真空吸附缺陷：弯曲变形延伸至书脊装订线，纸张在强行拉脱负压时发生撕裂。 - 垂直敞开复合通道：下部斜向吸入引导，上部垂直释放支撑，使纸页平稳摆动脱离真空区。 - Dany Qumsiyeh 在分析早期设计的失败原因时强调：“And basically the motion of the book is forcing the page to bend near the binding. And that's something that it can't do.” --- ### [09:43](https://www.youtube.com/watch?v=4JuoOaL11bw&t=583s) 传感器选型、机械打磨改版与书脊死角处理 在图像采集模组的选择上，Dany 直接展示了一根从普通馈纸式扫描仪中拆解出来的接触式图像传感器（CIS, Contact Image Sensor）。这类传感器内部集成了微型线阵光电二极管与脉冲彩色光源，通过依次闪烁红、绿、蓝光来捕获完整的全彩图像。项目团队拆解了大量消费级扫描仪，发现只要将原装的扫描棒与驱动电路板完整取出并重新连线，电脑操作系统就会将其直接识别为标准的 Canon 扫描仪，无需编写底层硬件驱动即可直接调用成像API，大幅缩省了开发周期。 ``` 棱柱顶部接触角示意（侧视）： ▲ 书籍装订处（书脊内槽） / \ / 9mm \ <-- 砂带机打磨后无法扫描的死区（原版因电路板更大） /| |\ / |CIS| \ <-- 矩形截面的接触式图像传感器（CIS） / |___| \ / \ 棱柱机身斜面 棱柱机身斜面 ``` 然而，将现成的 CIS 模块安装到棱柱形扫描架时遇到了严重的物理干涉：CIS 模块横截面呈长方形，无法完全塞进棱柱顶部的狭窄夹角深处，导致传感器扫描线与书籍内侧装订槽（Gutter/Groove）之间留有无法覆盖的死区。如果书籍排版留白很窄、文字紧贴书脊，边缘内容就会丢失。为了解决这个问题，Dany 采取了一种极为硬核的改装手段——直接将 CIS 传感器拿到砂带机（Belt sander）上打磨，将扫描棒边缘无用的电路板结构硬生生磨掉一截。令人惊喜的是，即使磨去相当一部分 PCB 边缘，传感器依然可以正常工作，从而将扫描死角成功缩小到距离装订线仅 9 毫米左右，足以满足绝大多数书籍的扫描需求。 现场观众提出是否可以通过增大棱柱顶角（如扩大到 90 度）来方便传感器贴近。Dany 解释称团队确实测试过该方案，但顶角过大会导致两重问题：其一，打开角度过大时书页无法自然平贴在斜面上，纸张本身的翘曲弧度反而会遮挡更深处的装订内容；其二，图书馆古籍保护专家普遍严禁将珍贵图书展开过大角度，扫描仪的开本角度越小对装订保护越有利。对于未来高标准的商用量产机，Dany 建议改用配备缩放透镜的线阵 CCD 传感器，通过光学投影将书脊拐角处直接成像，彻底摆脱 CIS 的物理外壳尺寸限制。在当前 300 DPI 分辨率下，原型机输出的扫描图像平整度与清晰度已与高档平板扫描仪完全一致。 - 图像传感器：采用拆解自常规扫描仪的 CIS（接触式图像传感器），利用 RGB 频闪合成色彩。 - 软件与电气复用：直接保留 Canon 扫描仪原装控制板，计算机免驱动免定制底层系统直接识别。 - 物理干涉难题：CIS 模块的矩形截面无法贴近锐角顶点，导致书脊内槽存在未扫描盲区。 - 硬件魔改：通过砂带机直接打磨 CIS 扫描棒电路板，将死角极限压缩至约 9 毫米。 - 顶角角度权衡：若加大顶角（如 90 度），书页受应力翘曲无法贴平，且图书馆专家对过度摊平书籍持强烈反对态度。 - 成像效果：当前默认以 300 DPI 运行，扫描平整度与图质比肩平板扫描仪；未来更优方案为采用透镜投影成像的线阵 CCD 传感器。 - Dany Qumsiyeh 在谈及手工打磨电路板的经历时风趣地提到：“And for this prototype, one way we improved that situation was I just took these sensors to a belt sander. And it turns out you can grind away quite a bit of that circuit board before the thing stops working.” ### [12:54](https://www.youtube.com/watch?v=4JuoOaL11bw&t=774s) 图像传感器优势与鞍座设计 ```mermaid flowchart TD accTitle: 鞍座与传感器系统交互 accDescr: 展示鞍座上的校准白条与夹具如何与扫描传感器交互并支撑书籍 subgraph SaddleAssembly["鞍座系统 (Saddle)"] Clamp["侧边夹具<br/>支撑书本重量/防止书脊弯曲"] WhiteStrips["校准白条<br/>原点位置覆盖传感器"] end subgraph SensorArray["扫描传感器"] CIS["接触式图像传感器 (CIS)<br/>单像素RGB全彩 / 300 DPI"] CalibProcess["开机原点校准<br/>识别基准白并自动调平"] end WhiteStrips -->|"原点覆盖"| CalibProcess CalibProcess -->|"基准设定"| CIS Clamp -->|"稳固固定"| SaddleAssembly ``` Dany Qumsiyeh 深入分析了选择线阵扫描传感器而非数码相机的核心技术考量。相机即使标称较高的百万像素，其每个物理像素点本质上也只是拜耳阵列中的红、绿或蓝单个通道，必须经过插值算法还原颜色；而扫描仪使用的接触式图像传感器（CIS）在 300 DPI 模式下，每一个物理像素都真实采集 RGB 三原色通道，呈现出更扎实、更高的图像质量。更关键的是成本差异——这类线阵传感器单价大约只需 20 美元，相较于高分辨率工业级相机而言极其低廉。 针对现场观众关于运动平稳性与扫描线性的提问，Dany 表示整个移动机构运行非常均匀，图像检查未发现明显的形变畸变。机器采用由步进电机与同步带（timing belt）构成的传动系统，这也是普通平板扫描仪的成熟方案。承载书本往复运动的滑台被称为“鞍座”（saddle）。鞍座两侧装有夹具，在初始归位（Home position）时，鞍座上的白色校准条会恰好覆盖传感器窗口，系统开机启动扫描前便以此确定基准白并完成色彩校准。 夹具的另一关键力学作用是支撑书本自重。扫描极厚的书籍时，如果扫描处于全书的最前端或末尾，绝大部分纸张都会堆积在一侧。若缺乏侧向夹具支撑，偏重的重心会导致书体倾斜下垂，使靠近装订线的纸张产生严重弯折，不仅破坏翻页气流环境，对于超大开本甚至会导致整本书从鞍座翻落。在回答观众关于装订类型的疑问时，Dany 指出这套装具兼顾精装书（hard-cover）与平装书（soft-cover），虽然精装书硬壳能与夹具配合得更加严丝合缝，但实际测试中平装书也能稳定运行。 - 传感器成本对比：约 20 美元的扫描传感器即可达到优于高价相机的色彩还原度。 - 鞍座夹具调整方式：将书本在中央放好后，手动将夹具推拢至微触书壳边缘即可，操作用时极短。 - 同步带与步进电机配合，确保了无须软件进行非线性位移补偿。 ### [15:55](https://www.youtube.com/watch?v=4JuoOaL11bw&t=955s) 控制系统与管道化软件架构 ```mermaid flowchart LR accTitle: 软硬件控制与数据流水线 accDescr: 展示MDrive电机通过串口换行符驱动SANE scanimage命令并由光电对射管闭环中断的流程 subgraph HardwareControl["硬件控制层"] Motor["MDrive 步进电机<br/>(内置微控制器)"] Sensor["对射式光电传感器<br/>(检测纸张到达)"] end subgraph SoftwarePipe["Linux 软件管道"] Serial["串口信号<br/>(周期换行符 \\n)"] Scanimage["SANE scanimage<br/>(--batch 模式)"] Viewer["Jeff 开发的开源图像浏览器<br/>(马赛克网格校验)"] end Motor -->|"单次行程起点"| Serial Serial -->|"管道传递 (Pipe)"| Scanimage Scanimage -->|"生成图像"| Viewer Sensor -->|"异常未遮挡/超时"| Motor ``` 扫描仪的驱动核心选用了一台集成微控制器的 MDrive 步进电机。运动控制逻辑极简，电机固件内预先烧录了速度参数与换向触发逻辑，一旦通电即可自主往复运行。电机通过串行接口与上位机相连，每当进入一个新往复循环的起点时，电机会通过串口输出一个换行符（newline），上位机借此获知鞍座已准时就位，充当极其精准的触发同步时钟。 软件架构充分体现了 Unix 管道 فلسفة 的极简主义。图像抓取直接依赖开源扫描库 SANE 的命令行工具 `scanimage`，调用其批处理模式（`--batch`），配置为“每次接收到回车键换行即执行一帧采集”。操作员只需在终端中通过管道将电机的串口输出重定向到 `scanimage`，整个采集循环就能完全自动化运转，底层完全无需手写额外的驱动通信代码。Dany Qumsiyeh 总结这一简洁设计时说道：“And so then all you have to do is take the line breaks from the motor and pipe them over to scanimage. And then the machine runs. So one command.”（所以接下来你所要做的就是获取电机输出的换行符，并通过管道传送给 scanimage。然后机器就跑起来了。仅仅一行命令。） 在图形界面层，主持人 Jeff 为项目编写了专用的开源图像查看器。该软件支持在扫描进行时实时瀑布流预览，并具有强大的“页码马赛克（Mosaic）”质检功能。用户只要点击任意页面底部的页码区域，系统就会自动将所有页面的该局部截取并纵向平铺排列。质检员顺着列向下扫视，看到“361、341、321、301”等页码连续等差递减，就能在几秒钟内确认扫描中途是否发生了漏页或双页重叠。 在硬件安全闭环方面，翻页槽内部布置了一道对射式光电传感器（break beam sensor），直接接入步进电机的微控制器 I/O 端口。每当单页纸张滑入吸附槽准备翻转时，纸张会瞬间打断光束。一旦在预定时间窗口内未检测到遮光信号（漏翻），或在错误的时序触发了遮光信号（卡纸或翻页不同步），电机固件会立即刹车停机，防止机械运动进一步扯碎纸张。 ### [18:48](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1128s) 原型迭代、生产率与现存局限 ```mermaid flowchart TD accTitle: 原型迭代历程与扫描吞吐分析 accDescr: 展示从废纸箱垃圾箱原型到金属机型的演进及性能指标 subgraph IterationHistory["原型演进"] P1["第一代原型<br/>办公室垃圾箱瓦楞纸板<br/>充气床垫电动气泵"] P2["第二代原型<br/>翻页几何形状优化试验"] P3["第三代原型 (现场展示)<br/>首台全金属加工机型<br/>BOM 成本约 1500 美元"] P1 --> P2 --> P3 end subgraph PerformanceMetrics["当前性能指标"] Setup["装夹准备: ~40 秒"] Speed["循环周期: 11 秒/双页<br/>(300页整书约30分钟)"] Scale["单人并发: 1人管理10台"] end P3 -.-> PerformanceMetrics ``` Dany 坦言现场展示的是该项目的“第一个非纸板金属原型机”，尚处于极早期的验证阶段。翻页扫描仪的起点极其粗糙：第一代原型完全是由办公室外垃圾箱里捡来的废弃瓦楞纸箱剪裁拼接而成，气动吸附源用的是一个打气垫床的小号电动充气泵。出人意料的是，这套零成本的纸板气动装置居然真的把书页翻动了。随后的两轮迭代优化了翻页槽的机械曲率，最终诞生了当前由数控加工金属件构成的第三代原型机。 在单机作业效能方面，该机器更换图书仅需约 40 秒来调谐夹具。往复周期为每对页面（2 页）耗时 11 秒，一本 300 页的标准书籍约需 30 分钟扫完。尽管绝对线速度不及工业级人工辅助翻页扫描仪，但其核心优势在于“全自动无人值守”。单名操作人员只需要在启动和结束时取放书本，理论上可一人同时看管 10 台设备并行作业。全机原材料成本（BOM）仅约 1,500 美元，其中近一半开销只是拆解用的一台市售平板扫描仪，其余主要是真空源和金属钣金件。 在现场问答环节，针对为何无法提升往复速度，Dany 直言瓶颈在于驱动电机功率不足。当鞍座装载大开本厚重书籍时，现有电机强行提速会发生丢步或动力不足，升级更大扭矩的电机或通过缩短机械行程均可立竿见影地提速。现场有观众问及纸张厚度与开本适用范围，Dany 透露，截至当时样机总共只测试了约 50 本书，在随机构建的书库中仅能成功处理约 60% 的书籍。其硬性物理限制在于尺寸两极化：超大书塞不进鞍座，超小开本（如袖珍口袋书）则因纸边宽度不足，根本无法触及气动真空吸槽。 更为严峻的可靠性挑战在于纸张损伤与漏页率。Dany 坦承在测试样本中，约有 45% 的书籍扫描过程出现过一到两页漏翻，同时约 45% 的书籍在某一页留下了折痕或撕裂。他特别澄清：“Of the books. Yes. Not the pages. So it's only a couple errors in the book. For most pages it was fine.”（指的是书本的占比，是的，不是纸张比例。所以整本书里往往只有两三个错误，绝大多数页面都是正常的。）屏幕上展示的破损惨剧，往往是因为纸张边缘在入机前已存在微小裂口，被高速气流及翻折槽刮擦后被进一步撕扯卷折。Dany 强调这类严重撕扯案例多数发生在光电对射传感器加装之前；安装传感器后，一旦进纸时序异常，机器便能即刻中断，避免损伤扩大。 - 纸张厚度耐受度：字典级超薄纸（Bible paper）容易被过强吸力扯坏，而过厚的封面硬卡纸则因刚度过大无法被气流吸起弯折，调节气泵负压是改善兼容性的重要旋钮。 - 材料缺陷与金属残渣：铝合金加工的翻页导板会因摩擦在书页边缘留下微量深灰色的氧化铝擦痕，未来机型亟需更换为食品级或医用级不锈钢材质。 - 视野死角（Gutter Loss）：现有 CIS 传感器封装存在物理边框，装订死角难以完整覆盖，后续需升级特殊的无边框 CCD 接触式组件。 - 终极构想（流水线联动机）：利用该机倾斜的进书槽几何空间，未来可以将数十台单机头尾无缝串联形成一条扫描长廊，两名工人在流水线两端一头进书、一头出书，实现爆炸性的数字化吞吐量。 ### [25:03](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1503s) 致谢团队与项目支持者 演讲接近尾声时，Dany Qumsiyeh 诚挚表达了对给予他支持的同事和管理层的感谢。他特别提到了同组工程师的关键贡献，指出对方为这个项目编写了全部的控制软件，协助传感器选型与调试，亲手安装了用于监测页面通过的遮光式光电传感器（break beam sensor），并帮忙拆解商用扫描仪提取零部件。更重要的是，在整个开发过程中，对方不仅持续提供技术建议，还在精神层面给予了坚定的支持，让他有动力将这个业余原型坚持做下去。 此外，Dany 感谢了他的现任主管 Stephane 以及前主管 Ram。他们充分理解并支持 Dany 利用 Google 的 20% 自由时间（20% time）开展这个硬件探索项目，并在资源与协调上给予了全力支持。致谢完毕后，演讲正式进入现场观众问答环节。 ### [25:41](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1541s) 故障处理机制与漏页检测 现场观众首先提问：当扫描过程中出现故障或异常时，机器会如何处理？Dany 解释了当前原型的错误响应机制。如果某张书页没有被真空吸附成功翻过，传感器便会侦测到这一状态，设备会立即制动停机。此时操作人员可以上前查看卡纸或翻页失败的具体原因，排除故障并重置系统后再次启动。 对于两张书页粘连同时翻过去（双页进纸）的情况，Dany 坦言目前的原型机尚无法检测到这种状况。一旦双页贴合通过，夹在中间的书页就会漏扫。针对这一痛点，他补充说明市面上的馈纸式扫描仪（sheet-fed scanners）早已具备成熟的双页检测方案（例如超声波或厚度传感技术），未来完全可以将类似的技术移植到线性扫描仪上以解决漏页问题。 ```mermaid flowchart TD accTitle: 线性书本扫描仪错误检测与处理流程 accDescr: 展示书页翻动失败停机以及未来双页粘连检测方案的逻辑图 subgraph ErrorHandling ["当前原型的异常处理机制"] A["书页翻动"] --> B{"遮光传感器检测<br>页面是否翻过？"} B -- "否（翻页失败）" --> C["立即停止机器运行"] C --> D["操作人员人工排查并重置"] B -- "是" --> E["继续正常扫描"] end subgraph DoublePageIssue ["双页粘连隐患与改进方案"] E --> F{"两页是否粘连？"} F -- "当前原型" --> G["无法检测，导致漏扫"] F -- "未来改进" --> H["引入馈纸扫描仪的<br>双页检测传感器"] end ``` ### [26:49](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1609s) 制造成本压缩与机身制作地点 随后有观众提问，考虑到扫描仪部件似乎是整套设备中最昂贵的部件，整机的成本是否还有进一步下探的空间。Dany 给出了极为肯定的回答，并算了一笔成本账：如果未来团队自行设计驱动电路板，成本将会急剧下降。机器中实际起成像作用的接触式图像传感器（CIS）硬件单价仅需 20 美元左右。如果实现规模化量产并自主配套电子器件，硬件成本会比现在直接拆解成品扫描仪便宜得多。 Dany 进一步指出，在实现量产时，最大的开销反而是组装工时费。机身本身的机械结构设计极其简单，目前所有的机架零件基本上都是平整的金属钣金薄板，激光或冲压切割非常容易且成本极低。紧接着有观众询问该原型机是在何处制造的，Dany 回答说全部是在 Google 园区内部的创客工坊（Google Workshops）加工制作完成的。 - 核心图像传感器硬件采购单价：约 20 美元/支。 - 机械结构特点：主要由扁平金属片构成，加工切割难度低、原料成本低。 - 主要潜在成本瓶颈：大规模量产时的组装工时与人工成本。 - 制造场所：Google 园区内部硬件工坊（Google Workshops）。 ### [27:36](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1656s) 脏污处理机制与真空吸尘的“意外特性” 一位观众提出了一个十分现实的问题：如果被扫描的书本非常脏，灰尘或污渍会不会直接蹭到扫描头表面，从而在整页扫描图像上拉出一条长长的贯穿划痕？Dany 承认确实存在这种物理风险，如果书页上附着了粘稠或糊状的污垢，确实会蹭在玻璃或传感器表面并留下一道长条伪影。要应对这一情况，系统需要配备相应的质检机制以及定期清理传感器的维护流程。 随后 Dany 笑着分享了一个团队内部常开的玩笑：“因为书本在整个运行过程中都是贴着真空吸气槽滑动的，所以机器在翻页的同时，其实顺便把书里的陈年灰尘都给吸走了。”他风趣地表示这甚至可以被算作该机器的一项“附加功能”（Feature）。 ### [28:12](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1692s) 书籍损伤风险与图书馆保护理念的权衡 有观众对比了现存的昂贵商用翻书机与该原型机，提出质疑：从机械结构来看，一旦发生机械故障，这台机器似乎更容易把脆弱的书页撕裂，而那些成熟的商用设备看起来则安全得多。Dany 坦承这种担忧，但他纠正道，即便是顶级的商用书本扫描仪也会发生撕页事故，只是概率极低，足够温和到能让图书馆放心使用的程度。 为了验证该方案的接受度，Dany 透露他们特意咨询了一位来自加州大学伯克利分校（UC Berkeley）的专业图书保护专家（conservationist）。令团队意外的是，这位专家对这台机器表现出了极高的热情。专家指出，在图书保护界，让公众获得全书数字化副本所带来的巨大知识传播收益，远远超过了个别书籍在极低概率下受损的微小风险。正如 Dany 转述图书馆的核心逻辑所言： > “And libraries accept the risk that, like when they check out books, they are going to come back in worse shape than they were before, but that it's worth it for people to use the book and to read the content.” 图书馆在向读者借阅实体书时，就已经预见并承担了书籍被还回时品相变差的磨损风险；只要能让读者获取其中的内容，这种损耗就是完全值得的。当然，Dany 也强调，团队后续会不断对机构进行改良，使其足够安全轻柔，以胜任那些纸质脆弱的古籍扫描任务。 ### [29:16](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1756s) 扫描传感器布局解析与未来展望 最后几位观众将焦点重新带回扫描头的物理布局上。有人询问为什么扫描传感器与底部的吸附切槽隔得相对较远。Dany 解释了这一布局的设计权衡：当前之所以放在那个位置，主要是为了让两根扫描棒（scan bars）正对排布，从而能够同一时间同步拍摄书页的正面与反面。此外，机器出书一侧的切槽实际上一直向上延伸到了扫描棒的所在位置。如果未来改变设计，将两根扫描棒拆分开来（例如左侧放一根、右侧放一根），那么扫描棒与切槽的间距就可以大幅缩短，从而缩短纸张必须走过的行程。 观众随即追问机器到底是由两台独立扫描仪拼成的，还是一台扫描仪带了两根传感器。Dany 证实这其实是一台完整的商用双面馈纸式扫描仪（sheet-fed scanner）拆解出来的。因为馈纸机本身就需要通过顶部和底部两根扫描棒同时扫描纸张的正反面，所以其内部原本就共享一块主板电路，团队只是把这两枚传感器拆下来固定到了设计好的机械槽位上。 当被问及这个 20% 时间项目未来的走向和商业化规划时，Dany 表示目前前途尚未完全确定，但活动结束后可以与感兴趣的同事进一步线下探讨。主持人 Jeff Breidenbach 随即打断并提议结束录像环节，现场在热烈的掌声中向 Dany Qumsiyeh 的分享致谢。 ## Notable quotes > [28:12](https://www.youtube.com/watch?v=4JuoOaL11bw&t=1692s) And libraries accept the risk that, like when they check out books, they are going to come back in worse shape than they were before, but that it's worth it for people to use the book and to read the content. — Dany Qumsiyeh ## Mentioned Dany Qumsiyeh, Jeff Breidenbach, Google, Google Books, Google TechTalks, Google Workshops, diybookscanner.org, Kirtas, Treventus, Canon, MDrive, SANE, scanimage, 加州大学伯克利分校 (UC Berkeley), Stephane, Ram ## Video description Google Tech Talk May 3, 2012 Presented by Dany Qumsiyeh ABSTRACT See a hardware prototype of an automatic, non-destructive book scanner. The machine turns pages automatically, and captures high-resolution images of each page. It was developed in 20% time, and built in the Google Workshops. There will be a live demo and a short technical presentation, describing the design process. Speaker Info: Dany Qumsiyeh is an engineer on the Google Books team. For more information on this project, see: http://code.google.com/p/linear-book-scanner/ [Source](https://www.youtube.com/watch?v=4JuoOaL11bw)

Source quotes

[28:12] Dany Qumsiyeh: And libraries accept the risk that, like when they check out books, they are going to come back in worse shape than they were before, but that it's worth it for people to use the book and to read the content.

Original transcript

[0:24] JEFF BREIDENBACH: My name is Jeff Breidenbach. [0:25] I'm very pleased to introduce Dany Qumsiyeh, who is going to [0:29] be presenting on an automatic book scanner. [0:32] This talk began with an idea of showing this device to a [0:36] single person. [0:38] And that person invited two more people. [0:40] And you're all here through the [0:41] incredible power of recursion. [0:44] So with that, Dany, come take it away. [0:50] DANY QUMSIYEH: Thanks, Jeff. [0:52] Let me take a moment and shut it down. [1:04] I'm just going to tell you about this book [1:05] scanner that I built. [1:08] But first I want to just show you a few examples of some [1:11] book scanning technologies that already exist for [1:14] comparison. [1:15] The easiest way to scan a book, of course, is to cut off [1:17] the binding and then stick it through a sheet-fed machine. [1:20] There are machines that will just feed each sheet through [1:22] and scan it. [1:24] Very high quality, but the problem is [1:25] you destroy the book. [1:27] So what we're talking about here are [1:28] non-destructive book scanners. [1:32] So there's a large class of book scanners that are not [1:35] totally automatic, where a person turns each page. [1:39] And a lot of the time, there are glass plates that flatten [1:42] the pages out. [1:43] And then there are cameras that take a [1:44] picture of both pages. [1:46] So you have to kind of raise this glass plate, turn the [1:49] page, lower the plate, and then snap a picture. [1:53] So it's pretty fast, but involves a [1:56] lot of manual labor. [1:59] There's actually a large community of people at the [2:01] diybookscanner.org that build their own. [2:03] And they post instructions for how to build your own. [2:06] So they make them out of cardboard and wood and all [2:08] kinds of things that are lying around. [2:12] There are some commercial robotic book scanners. [2:17] There's quite a few that use some sort of robot arm the [2:22] same way that a person would turn a page. [2:24] And Kirtas is one of these examples. [2:30] They have an arm that kind of suctions a page and then moves [2:32] it over to the other side. [2:39] There's another that uses a very different method. [2:42] This company is called Treventus. [2:45] And they make a machine that has this V-shaped thing that [2:48] moves up and down. [2:49] And it actually suctions two pages against it while [2:52] capturing images with a linear sensor along the tip. [2:57] It's really cool. [2:58] And these are state of the art machines. [3:01] And they've gotten a lot of praise for [3:02] being gentle to books. [3:06] The main issue is that they're really expensive. [3:10] One of these Treventus machines can cost on the order [3:13] of $100,000. [3:19] This is the machine that I built. [3:20] And I'm going to show you a brief video so that you can [3:24] see some closer up views of it. [3:34] And then I can go into some of the technical details. [3:49] The general idea is this book is moving back and forth on [3:52] top of this prism-shaped machine. [3:54] There are linear sensors, just like in a flatbed scanner that [3:59] the book moves across and it captures an [4:01] image of each page. [4:03] And then there's a vacuum, in this case just a regular house [4:06] vacuum, that's connected to those spots [4:08] that's sucking air. [4:09] And it sucks the single page through and that page goes [4:13] across to the other side. [4:26] AUDIENCE: [INAUDIBLE] [4:31] DANY QUMSIYEH: The question was how do you ensure that you [4:32] only get one page? [4:33] That's the next thing I'll talk about once this is done. [4:40] And then the images show up on a computer [INAUDIBLE]. [4:51] So it's really simple. [4:52] And it doesn't involve very many parts, so it's not very [4:57] expensive to build. [5:02] The question was, how do you ensure that it [5:03] only turns one page? [5:05] I'm going to try and provide a very simple diagram of what's [5:09] going on when a page gets peeled off. [5:12] In this case, we have a stack of pages that are sliding in [5:15] from the right and I'm looking at them edge on. [5:18] And they're sliding along this curved surface. [5:20] And along that curve, there are slots where I'm applying [5:25] vacuum suction. [5:27] Hopefully it's obvious that one page gets [5:30] sucked against this. [5:31] The other pages don't actually receive any suction, so they [5:34] will tend to stay straight and go across. [5:38] And then they get divided by this piece of [5:39] metal on this side. [5:42] Now, it's still possible for two pages to [5:43] kind of be stuck together. [5:46] And to help out with that, in this prototype there's [5:49] actually air that blows up from underneath onto the edges [5:53] of the pages. [5:54] And that helps to separate pages if [5:55] they're stuck together. [6:00] Next I want to talk about why the channel that the page goes [6:05] through looks so strange, why it's shaped the way that it [6:08] is, and try to explain the motivation for that. [6:12] And I'll try to explain it by explaining a few examples that [6:14] don't work. [6:16] So for example, if you have this prism shape-- [6:20] and I'm looking at a top view here-- [6:22] where a page on one side has to get from that side over to [6:25] the other side, the simplest thing to do is to cut a slot, [6:30] like a diagonal slot in that prism so that the page can [6:35] slide into it and go across to the other side. [6:38] The problem you run into is when the page is in the middle [6:41] of that slot it actually has to kind of twist. [6:43] So the binding of the book stays straight, but the page [6:46] has to twist along that slot. [6:48] And that's something that it just can't do, so it's very [6:51] likely for the page to tear. [7:00] So the way that the machine is really shaped is a little bit [7:03] more like this, where there's a flat section in the middle [7:06] of that diagonal slot. [7:08] The idea is the page is coming in. [7:10] And while the page is in that intermediate section, it's [7:13] actually free to swing across to the other side. [7:16] And so, it can go to the other side without actually getting [7:18] twisted, so that works a lot better. [7:25] There's another problem you run into. [7:27] In this example, the slot for the channel is just a straight [7:32] diagonal along that front surface. [7:35] And there's another problem you run into if you have that [7:37] straight diagonal and you're applying vacuum pressure to [7:40] try and pull the page through. [7:43] This is trying to illustrate that. [7:45] Basically this is a side view. [7:46] And here's the page coming from the right to the left. [7:49] And the darker gray area is the channel where the page is [7:54] sliding into. [7:55] And that really dark slot is where the [7:57] vacuum is being applied. [7:58] That's where the slots of the vacuum are. [8:01] In this case, the page is coming from the right. [8:03] And the corner of it is getting sucked up against [8:05] those slots and curving around that curve. [8:10] So as the page continues, you run into a problem where the [8:15] page is being curved along this diagonal and that curve [8:19] reaches the binding. [8:20] And basically the motion of the book is forcing the page [8:25] to bend near the binding. [8:26] And that's something that it can't do. [8:28] Basically, in order for the page to continue to the left, [8:30] it has to pull itself off of that vacuum area. [8:34] And if that requires too much force, then the [8:36] page will just tear. [8:39] So this doesn't work very well. [8:41] The picture at the top was actually an earlier cardboard [8:44] prototype that looked a little bit more like this. [8:46] And that was the problem that it had. [8:50] So the way the machine is actually shaped is a [8:52] bit more like this. [8:54] There's a diagonal opening near the bottom. [8:57] And then the opening opens up vertically in the top part. [9:02] So as the book is going across, the page gets [9:06] suctioned against that dark area and [9:10] curves around that diagonal. [9:13] But at some point, the page reaches the part where this [9:17] opening opens up vertically. [9:20] And so a page is no longer supported by the surface [9:22] that's on the right. [9:23] So it's actually free to swing across to the other side. [9:27] So if you have a shape like this, then the page can [9:32] basically slide its way off of that vacuum section without [9:36] applying too much force to the binding. [9:43] So the sensors-- [9:44] let's see. [9:45] I've got one here. [9:47] So these are the sensors. [9:49] It's a contact imaging sensor. [9:51] And it's straight out of a flatbed scanner [9:55] or a sheet-fed scanner. [9:56] It's the same thing that they use. [9:58] It has an array of light sensors and pulses colors so [10:01] that you can capture each color. [10:04] And it actually did come from a sheet-fed scanner. [10:07] We took it apart-- we took apart a [10:09] lot of scanners actually. [10:11] And it turns out if you take out the circuitry and these [10:15] scan bars and just plug them together, you can use [10:18] it just like that. [10:19] You can tell the computer to scan and it thinks the scanner [10:22] is just a regular Canon scanner. [10:25] And it captures images. [10:27] So there's very little software required. [10:30] And we didn't have to do any of the electronics. [10:34] One note about using these sensors is that the sensor is [10:38] kind of square. [10:40] And when you stick it inside-- [10:42] this is an end view of the machine with a [10:44] book open on top. [10:45] When you stick it inside, you can't actually stick it all [10:48] the way up into that corner. [10:50] And that can present a problem, because you can't [10:52] capture an image all the way into the groove of the book. [10:55] So if the book has very narrow margins or has content that's [10:58] all the way into the groove, then you might not be able to [11:00] see that content. [11:02] And for this prototype, one way we improved that situation [11:07] was I just took these sensors to a belt sander. [11:11] And it turns out you can grind away quite a bit of that [11:14] circuit board before the thing stops working. [11:19] And that was enough to get us within about nine millimeters [11:24] of the groove, which is OK for most books. [11:30] For a real machine if you want to do a lot of scanning, it's [11:34] probably better to use a different sensor. [11:37] There's a lot of other scanners that use a CCD [11:39] sensor, where they have a lens and they project the image [11:42] onto a linear sensor. [11:44] And in that case, you could actually project the image all [11:46] the way from the corner. [11:48] It wouldn't have this kind of square restriction. [11:50] AUDIENCE: [INAUDIBLE] [11:53] DANY QUMSIYEH: Yes. [11:54] Somebody mentioned that increasing the angle of the [11:59] corner at the top would help. [12:02] And that's something we looked into a bit. [12:06] One problem you run into when you make that angle larger, [12:09] like 90 degrees, is that the book doesn't actually want to [12:12] sit completely flat. [12:13] So the curve of the page may actually still hide some of [12:19] the binding from you, because it's not totally flattened out [12:22] onto the surface. [12:25] Also, librarians generally like to keep the-- [12:31] they consider it very bad to open the book too far. [12:33] And so if you can keep the angle of opening as small as [12:37] possible, that's better for the book. [12:42] This is the result you get. [12:43] It's a very flat scan, as if you had put the page onto a [12:47] flatbed scanner. [12:48] And the quality is exactly the same. [12:50] So by default, I'm running this at 300 DPI. [12:54] But you could crank it up larger if you want. [12:58] And it's also a lot better than using cameras for [13:01] scanning, because cameras, even though they may have a [13:06] certain number of megapixels, each pixel is actually red, [13:09] green, or blue. [13:10] And in these scanners, when you say 300 DPI, you're [13:12] actually capturing all three colors for each pixel. [13:18] So it works a lot better. [13:20] It has much higher quality. [13:22] And these sensors are only like $20, compared to a very [13:26] expensive camera that you would need. [13:28] Yes? [13:30] AUDIENCE: [INAUDIBLE] [13:33] DANY QUMSIYEH: The question was, how does the linearity of [13:35] motion look? [13:36] Do I have to do any correction for the motion [13:38] not being very constant? [13:40] And it seems pretty good. [13:42] We've looked at the images and haven't noticed any [13:46] disturbances. [13:46] The motor is a stepper motor and it has a [13:49] very controlled motion. [13:50] And the belt that we're using is a timing belt, the same as [13:53] in a regular scanner. [13:56] Earlier, we had experimented with ways of correcting for [13:58] that using kind of a pattern that's on this saddle. [14:06] I'll say a little bit about this, we're calling it a [14:09] saddle, the thing that moves the book back and forth. [14:11] It has clamps on it to kind of hold the book in place. [14:16] And another purpose of the saddle is it has these white [14:19] strips on it which are used for calibrating the sensors. [14:22] So when it's in the home position, those strips cover [14:25] up the sensors. [14:27] So when you first start scanning, they figure out what [14:30] color is white and they calibrate to that. [14:34] The other important function of these clamps is to help [14:37] support the weight of the book. [14:40] If you don't have them, you run into a problem where, say [14:43] you have a very thick book and you're at the very beginning [14:45] or end of the book, so all the pages are to one side. [14:48] That weight can cause the book to kind of sag. [14:51] And it causes the pages to sort of [14:52] curve near the binding. [14:55] And that can cause problems for page turning. [14:57] Or the book could just fall off to one side [14:59] if it's very large. [15:02] Yes? [15:02] AUDIENCE: [INAUDIBLE] [15:07] DANY QUMSIYEH: The question was, does this mostly work for [15:10] hard-bound books, or does it work for [15:11] soft-cover books as well? [15:13] And it does work for soft-cover books. [15:16] I guess the support from the clamps kind of works a little [15:22] bit better for hardcover books. [15:24] But we haven't run into any problems using it on [15:27] soft-cover books. [15:31] Yes. [15:32] The clamps on the saddle look kind of like that. [15:36] And they help support the weight of the book. [15:38] Basically, when you first start scanning, you set the [15:41] book down somewhere in the middle. [15:43] And then you adjust these so that they're just [15:45] touching the book. [15:47] And then from then on, if the book is to one side or the [15:49] other, it helps to support the weight. [15:55] The motor we're using I said is a stepper motor. [15:57] And it's an MDrive. [15:59] And we have this timing belt that goes around the top and [16:01] then down through the machine. [16:04] And this motor actually comes with a built-in [16:07] microcontroller, so pretty much the entire code for it is [16:12] on the right. [16:13] Basically, you preprogram it with the code that tells it [16:16] how fast to go and when to change direction. [16:20] And from then on, when you power it up, it just goes back [16:24] and forth and does its thing. [16:26] The extra thing that it does is it has a serial connection. [16:30] And at the beginning of each cycle, it outputs a new line. [16:35] And we use that for timing the actual scan, so that we know [16:38] when the motor's arrived. [16:42] The software for this is really simple. [16:44] The scanimage program that shows there, that's from an [16:49] open source scanning library called SANE. [16:52] And so we didn't need to write any software for that. [17:00] And we call that with a flag that tells it to capture an [17:03] image every time you hit Enter. [17:05] That's the batch prompt. [17:07] And so then all you have to do is take the line breaks from [17:12] the motor and pipe them over to scanimage. [17:17] And then the machine runs. [17:18] So one command. [17:25] There's some extra software that Jeff, who [17:27] introduced me, wrote. [17:30] This is just an image viewer. [17:32] It shows you the images as they're being scanned. [17:34] And it's open source, the URL at the top. [17:38] And it has a lot of great features, like you can click [17:41] on a page number and you can see a mosaic of all the page [17:44] numbers that you've scanned. [17:46] And that makes it really easy to tell whether something went [17:49] wrong during scanning, like you missed a page or you [17:51] duplicated a page. [17:52] Basically, you can look down these columns and see like [17:55] 361, 341, 321, 301. [17:58] You can see all the way down that everything went fine and [18:02] that no pages were missing. [18:10] Another way that we avoid problems and detect duplicate [18:14] pages or missing pages is with a sensor that's right inside [18:19] the machine. [18:20] As soon as a page comes in the slot to be turned, there's a [18:24] break beam sensor. [18:25] So we detect exactly when the page arrives. [18:28] And basically, if a page arrives at the wrong time or [18:32] doesn't show up, then we know that something's wrong and the [18:35] motor stops. [18:36] So this sensor is connected directly to the motor. [18:38] And that's part of that microcontroller program. [18:48] That's about all the details. [18:50] Just to give an idea of where things are, this is the first [18:54] non-cardboard prototype. [18:56] So there's still a lot to improve. [19:00] The first prototype on the top left, that's cardboard just [19:03] from a dumpster that was outside the office. [19:05] So it cost nothing. [19:07] And it had a little air pump from an inflatable mattress. [19:12] And that was actually able to turn pages. [19:14] And it was really exciting. [19:18] I experimented with the shape of the page-turning mechanism [19:21] a few times, but this is the first model that's actually [19:26] built out of metal. [19:30] Where it is right now, basically it's very quick to [19:35] set up for a book. [19:36] It takes about 40 seconds to adjust those clamps and start [19:40] the scanning. [19:42] The cycle time of this going back and forth right now is 11 [19:45] seconds for a page pair. [19:47] So a 300-page book would take about half an hour. [19:51] This is a lot slower than many of the alternatives. [19:56] But since it's fully automatic, one operator should [19:59] be able to run like 10 of them. [20:02] You don't have to do more than just start it and stop it. [20:07] The materials, there's only about $1,500 of materials. [20:11] About half of that is just the scanner that we took apart. [20:14] And there's the vacuum and then the metal. [20:17] So it should hopefully be very cheap to build. [20:24] Yes? [20:24] AUDIENCE: [INAUDIBLE] [20:32] DANY QUMSIYEH: The question was, it is possible to get any [20:33] sensor feedback from the motor as to whether the page is hung [20:37] up or tearing? [20:40] Yes. [20:40] There are models of this motor where it will know if it's [20:42] stalled, if something is stuck. [20:45] And I'm sure there are models where you could also detect [20:47] how much force it's taking to move. [20:51] So yes, if you wanted to get a fancier motor, [20:53] you could do that. [20:54] Yes? [20:54] AUDIENCE: [INAUDIBLE] [20:57] DANY QUMSIYEH: What prevents me from running it faster? [20:59] Right now that's actually the motor. [21:01] If I try to run it faster for a large book, then the motor [21:03] doesn't have enough power to move it. [21:05] So an easy way to make this faster is just to [21:08] get a better motor. [21:10] There's also things you can do with the arrangement of the [21:11] sensors and the shape of the thing. [21:13] Basically just make it so that it has a [21:15] shorter distance to travel. [21:19] AUDIENCE: [INAUDIBLE] [21:28] DANY QUMSIYEH: The question is, do you need any [21:29] adjustments for book size or paper thickness? [21:34] There's a limit to the size of books that this can scan. [21:37] And right now, I guess I was going to mention we've only [21:40] scanned about 50 books. [21:43] Basically about 60% of the books that we've come across [21:47] are able to be scanned. [21:48] Many of them are too big or too small. [21:51] So basically, it has to fit inside of the saddle. [21:53] And it also has to be able to reach those vacuum slots. [21:55] If a book is really small then it won't reach them at all. [22:00] And so that's a number that could also be improved by [22:03] changing some of the dimensions, basically doing a [22:06] bit more experimentation with how the thing is shaped. [22:10] And then for thin pages or thick pages, one major factor [22:16] in that is what the power of the vacuum is. [22:19] So it's best if you can adjust that. [22:24] For very thin pages you don't want to pull too hard on them. [22:28] But again, there's a limit on what you can do for that. [22:31] So this works on most books we've come across. [22:33] But it won't work on like dictionary-thin pages or [22:37] really stiff kind of cover page material. [22:44] We still have a lot of problems. [22:46] Basically about 45% of books have one or two skipped pages [22:51] that we tried. [22:53] And about 45% of them have some sort of fold [22:58] or tear in the page. [23:00] That's really bad. [23:00] This image is an example of one of the bad cases. [23:05] And actually what happened here was the page already had [23:08] a little bit of a tear in the corner. [23:10] And that presents a problem, so when it went through the [23:12] machine, it kind of exacerbated [23:14] that and folded it. [23:15] AUDIENCE: [INAUDIBLE] [23:17] DANY QUMSIYEH: Of the books. [23:18] Yes. [23:18] Not the pages. [23:19] So it's only a couple errors in the book. [23:23] For most pages it was fine. [23:25] AUDIENCE: [INAUDIBLE] [23:32] DANY QUMSIYEH: We'll leave those questions until [23:33] afterwards, if that's all right. [23:38] And also, this was before we installed the sensor. [23:41] So in this case where the page was folded it's very likely [23:45] that the sensor would have detected that the page didn't [23:47] arrive at exactly the right time and stopped the machine [23:50] before it caused any more damage. [23:54] There's a lot of improvements to make. [23:57] I still haven't experimented much with different shapes. [23:59] This is only the third iteration, really. [24:02] And there's adjustments you can make to the vacuum, like [24:06] maybe adjusting where the vacuum is applied and what the [24:08] slots look like. [24:09] There's better materials, like aluminum leaves a little bit [24:12] of aluminum oxide residue that you can kind of see on the [24:15] edge of the book. [24:17] And it would be better if it were stainless steel or [24:19] something like that. [24:20] And of course, better image sensors like a CCD sensor that [24:23] would be able to capture all the way into the corner. [24:27] And then there are crazier ideas, like you could take a [24:29] whole bunch of these and chain them up if you have a lot of [24:33] books to scan. [24:34] Basically, this diagonal opening can be chained pretty [24:38] close together. [24:39] So you could have just two people, one at each end, [24:44] loading books one right after another. [24:46] And you could perhaps scan books really quickly that way. [24:51] There's a lot of interesting ideas about [24:53] what to do with it. [24:56] So that's it. [24:58] I just want to be sure to thank Jeff, the person who [25:02] introduced me. [25:03] He's been really crucial to this product. [25:07] He wrote all the software. [25:09] He helped out with the sensors. [25:10] He installed the break beam sensor and [25:12] helped take apart scanners. [25:14] And he's given me advice all along the way and given me the [25:19] moral support to keep working on it. [25:22] And I also want to thank my manager, Stephane and [25:25] previously Ram, who have been OK with me working on this. [25:31] And they've supported me in any way that they could. [25:37] That's about it. [25:39] I can take questions. [25:41] AUDIENCE: [INAUDIBLE] [25:47] DANY QUMSIYEH: The question was, can you talk a little bit [25:48] about error [INAUDIBLE] [25:50] and what you do when something goes wrong? [25:53] So right now with this prototype, basically if a page [25:56] doesn't turn, that detects it and the machine just stops. [26:01] And so you can notice that and come and figure out what [26:04] happened and then reset it and start it again. [26:10] Right now, there's no way to detect if [26:12] two pages go through. [26:13] If they're really stuck together, then that means that [26:16] you missed capturing the images of the [26:18] page that was in between. [26:21] That's a problem, but there are ways that you can detect [26:23] that and there are things like sheet-fed scanners used to [26:29] detect whether two pages have gone through. [26:32] And you could do something similar. [26:38] Yes? [26:38] AUDIENCE: [INAUDIBLE] [26:49] DANY QUMSIYEH: The question was, do you think you can make [26:50] it any cheaper, because it sounded like the scanner part [26:54] was already the most expensive part. [26:57] Yes. [26:57] If you did your own electronics, for example. [27:00] These sensors, the part that we're actually using [27:02] are only $20 each. [27:04] So if you built the electronics yourself at a [27:07] large scale, then it would probably be even [27:09] cheaper than it is now. [27:11] But the biggest cost will probably just [27:13] be putting it together. [27:16] And right now, all the pieces are basically [27:19] flat pieces of metal. [27:21] And so they're very easy to cut. [27:24] Any other questions? [27:30] Yes? [27:30] AUDIENCE: [INAUDIBLE] [27:32] The question was where did I build it. [27:34] And I built on some workshops that are on campus. [27:36] AUDIENCE: [INAUDIBLE] [27:41] DANY QUMSIYEH: The question was what happens if the book [27:42] is really dirty? [27:43] Does it leave a streak all the way down the page? [27:46] If there's really something like smeared on the paper, [27:48] then yes, that could get stuck to the [27:50] sensor and leave a streak. [27:52] And you'd probably have to have some process to notice [27:55] that and clean it up. [28:00] But we joke that since the book is travelling over these [28:04] vacuum slots, we're actually cleaning out dirt from the [28:06] book as it goes. [28:07] And that's a feature. [28:11] Anyone else? [28:12] AUDIENCE: [INAUDIBLE] [28:17] DANY QUMSIYEH: The question was, it looks like this [28:19] machine, if something goes wrong, could [28:21] actually tear a page. [28:22] And it looked like the other machines we looked at couldn't [28:25] actually tear a page. [28:27] The other machines, they do tear pages. [28:29] It's just very rare. [28:32] And they're gentle enough that libraries are [28:35] OK with using them. [28:36] For this machine, we've actually talked to a [28:39] conservationist from Berkeley. [28:41] And he was really excited basically that the benefit of [28:48] having digital versions of these for people to use was [28:50] greater than the small risk of causing damage to some books. [28:55] And libraries accept the risk that, like when they check out [28:58] books, they are going to come back in worse shape than they [29:02] were before, but that it's worth it for people to use the [29:05] book and to read the content. [29:09] But yes, hopefully this can be improved to the point that is [29:13] safe enough to use on fragile books. [29:16] Yes? [29:17] AUDIENCE: [INAUDIBLE] [29:22] DANY QUMSIYEH: The question was, why is the scanner so far [29:23] away from the slot. [29:25] The main reason it's there was just we wanted to have both [29:28] scan bars together so that they capture at the same time. [29:32] And the slot on this side, the exit slot actually extends all [29:36] the way up to where the scanner is. [29:39] So you could actually make them closer if you had them [29:41] separate, like one on this side and one [29:43] on the other side. [29:45] And that might make a shorter distance to travel. [29:47] AUDIENCE: [INAUDIBLE] [29:54] DANY QUMSIYEH: The question was, is it two scanners or [29:55] just one scanner with two bars? [29:56] This is one scanner with two. [29:58] It's from a sheet-fed machine, so the sheet would feed [30:00] through and it had a scan bar on the top and on the bottom. [30:03] And so it was one circuit board and we [30:04] just took two sensors. [30:09] Any other questions? [30:10] AUDIENCE: [INAUDIBLE] [30:13] DANY QUMSIYEH: The question is, where does this [30:14] project go from here? [30:15] And I don't know. [30:16] But we'll talk a little bit more about that later. [30:19] AUDIENCE: [INAUDIBLE] [30:21] JEFF BREIDENBACH: Maybe this is a good time to end the [30:22] recorded portion of the talk. [30:25] And let's give Dany a hand. [30:28] [APPLAUSE]