Document scanners: More complicated than "Just get the Fujitsu"?

TidBITS Talk
61

This is interesting to me in particular because I’m currently writing a receipt organizer app. I have not seen evidence of what you’re saying, but perhaps I’m misunderstanding, or have a slightly different case. If I feed my app an upside down receipt from the local grocery store, Apple’s Vision framework OCR’s it just as well as it does all the other correctly oriented receipts from the same store. I wonder what the difference could be? Perhaps locale?

1 Like
62

You’re right, and my post was broader than my evidence — I wrote “Apple’s Vision text recognition” when what I had actually measured was one recognition level. I went back and tested it properly.

Two Spanish documents: a crumpled Transbank receipt from a café here in Santiago, and a full page of body text from an official form. macOS 26.6.2. Real pixel rotation, handler told .up, usesLanguageCorrection off in both cases, scored by how many recognised words survive a spellcheck.

Full page, 2200 px:

  • .accurate — 180/182 real words upright, 180/182 inverted. It simply reads it.
  • .fast — 168/180 upright, 67/142 inverted. Character counts: 2021 upright, 2022 inverted.

Receipt:

  • .accurate — 26/26 both ways. Confidence was slightly higher inverted.
  • .fast — 21/22 upright, 11/23 inverted.

So on .accurate, which for a receipt organizer you almost certainly are, upside down reads exactly as well as right side up, and you were right to push back. The “JE6Ed E |EIOI” I quoted came from .fast.

Why I was there at all: I’m deciding orientation, not extracting text, so recognition runs four times per image (0/90/180/270) before anything else happens. Four .accurate passes is too slow for that, so the decision runs on .fast and the real extraction happens afterwards on .accurate.

Which makes the honest version of my claim narrower, and I think still worth having: under .fast, the two orientations differ by one character out of two thousand. Volume of text cannot tell them apart. Whether the words are real can, and that part holds.

On locale — I can’t rule it out, since everything I tested is Spanish and yours is English. But the gap between the two levels is far larger than I’d expect a language effect to be, and it’s one line to check on your side.

One thing that may be worth your time regardless: if any part of your pipeline uses .fast for a cheap first pass, an upside-down receipt hands you the same volume of text with the wrong content. It fails quietly rather than returning less.

63

Thanks @sdiazthomas. I’m sure the instructions/options on the recognizeTextRequest are the difference. I only use recognitionLevel = .accurate, and usesLanguageCorrection = true. It looks like you’re being nicer to your users than I am, and attempting to auto-rotate before recording. I just give it back to them the way I got it (since the OCR works regardless), and give them a rotate button for the image if they care to do so.

Happy to discuss further if there’s any need, but I’d better not wander too far off-topic here. :slight_smile:

64

In researching the currently available document scanners I’ve learned that Ricoh and Epson are the principle players and their top models share many of the same capabilities and features. The models I’m looking at are even the same price.

That said, there are several aspects that do separate them and I can’t settle on which warrant serious consideration or can be ignored.

The differentiations include;

Epson ES-550W (similar to ES-590W minus touchscreen)
– Epson ScanSmart AI software –for AI-ready data
– PDF (presumably searchable), Word, Excel outputs
– WiFi is 802.11 b/g/n WPA3 (2.4 GHz only)
– USB is SuperSpeed 3.0
– Feed handles mixed sizes with skew correction, staple protection
– Auto cropping and blank page removal
– TWAIN compatible
– 600 dpi, 30-bit input
– no touchscreen, physical buttons instead
– Documentation in downloadable PDFs
– a Costco specific model (no-hassle return if needed)

Ricoh iX2500
– ScanSnap Home software, Kofax Power PDF, ABBYY FineReader OCR
– PDF, JPG, searchable PDF outputs
– WiFi is 802.11 b/g/n/ac/ax (WiFi 6) WPA3 for both 2.4 and 5 GHz bands
– USB is Type-C USB 3.2 Gen1x1
– multi-feed detection, skew detection. Separate receipt feeder attachment reduces capacity by half
– no TWAIN support
– 600 dpi, ? bit depth input
– 5 in. touchscreen, does allow standalone operation
– Documentation principally web-based

Which, if any, of the above are deal breakers or crucial advantages, and why?

1 Like
65

I’d research longevity. Especially with respect to software. Any scanner that isn’t total junk will outlive the manufacturer’s software support. So look at what third-party packages may also support it. Look over the web site for VueScan and maybe other packages.

My comments:

  • When asking about products, always share links, so we can see what you’re considering. Here are the ones for yours:
  • These are page scanners. Nothing wrong with that, but you won’t be able to scan large objects like books. On the other hand, a flatbed scanner can be really inconvenient if you need to scan a stack of papers.
  • Software:
    • Nothing bundled with the scanner will be high-end, but people do seem to like ScanSnap.
    • ABBYY FineReader used to be bundled with everything. I don’t know how well it works, but it must be capable if they’re still selling it after all these years.
    • VueScan says it supports both the ES-550W and the iX2500 without any additional software. Which means you will have software support, effectively for the life of the hardware, no matter what the manufacturer chooses to do.
      • If you aren’t sure if it’s worth the price, you can download a trial copy. It will put a watermark on your scans, but it should be good enough to determine if it’s worth paying for.
    • I couldn’t determine if either company offers an ICA driver for Apple’s Image Capture tool. But maybe that is distributed via Aple.
  • It’s weird, but neither product’s manuals mention the USB interface beyond pointing out that it exists. It seems that they really really want you to do all your scanning via Wi-Fi
  • Wi-Fi connectivity
    • Epson
      • You can configure the device via their app or via a web page.
        The manual tells you how to determine its IP address to use the web interface
        No Ethernet interface. Wi-Fi and USB connectivity only
      • The manual says you can scan directly to a smart device or cloud service. Not sure what this means in practice.
    • Ricoh
      • Their manual says it can scan directly to an SMB file share, like a Mac’s shared folder or a NAS. That can be convenient.
      • It looks like they really want you to create a ScanSnap account and use it for managing all your scans/receipts. The manual focuses almost entirely on that.
  • TWAIN is irrelevant. Apple hasn’t supported it for some time.
4 Likes