Optional list of images to ignore when inferring nonvisual reading

[HadrienGardeur]

Our current inference rule for "accessModeSufficient": "textual" is limited to publications that have either:

• zero images
• or where the only image present is a cover

We'd like to add the ability to ignore images based on their cryptographic or perceptual hash as well. This would be used for a list of well-known images (such as logos or decorative images shared across a collection) that would be automatically considered as decorative when inferring "accessModeSufficient": "textual".

We need to support two types of inputs:

• a directory of images to ignore (for a manual workload)
• or a list of hashes (for automated workloads)

[chocolatkey]

@HadrienGardeur Based on what I've done so far for inferring perceptual hashes, this is probably not something we'll be able to do in realtime, meaning use for streaming purposes, as the calculation of the perceptual hash of all images in an EPUB could take significant time.

[HadrienGardeur]

@chocolatkey that's fine, we don't necessarily have to expose the same options when using the manifest or the serve commands.

[chocolatkey]

@HadrienGardeur Opinion on the added arguments?

Usage:
  rwp manifest <pub-path> [flags]

Flags:
  -h, --help                                  help for manifest
  -i, --indent string                         Indentation used to pretty-print
      --infer-a11y string                     Infer accessibility metadata: no, merged, split (default "no")
      --infer-ignore-image-directory string   Ignore the images in a given directory when inferring textual accessibility.
      --infer-ignore-image-hashes string      Ignore the given hashes when inferring textual accessibility. Hashes are in the format <algorithm>:<base64 value>, separated by commas.
      --infer-page-count                      Infer the number of pages from the generated position list.

[mickael-menu]

I know it's starting to be a mouthful but for consistency I would keep the infer-a11y prefix for these new options:

      --infer-a11y string                      Infer accessibility metadata: no, merged, split (default "no")
      --infer-a11y-ignore-image-dir string     Ignore the images in a given directory when inferring textual accessibility.
      --infer-a11y-ignore-image-hashes string  Ignore the given hashes when inferring textual accessibility. Hashes are in the format <algorithm>:<base64 value>, separated by commas.

[chocolatkey]

OK, will adjust that

[HadrienGardeur]

I know it's starting to be a mouthful but for consistency I would keep the infer-a11y prefix for these new options:

We can also use a single letter or two letter shorthand, like we do for help and indentation.

[chocolatkey]

@HadrienGardeur Something we may want to consider adding is support for in the future is ignoring gaiji images, which will be common in Japanese publications.

[HadrienGardeur]

@chocolatkey this will require additional CSS parsing to detect the use of an image to replace a character, unless there's another way that I'm unaware of.

[chocolatkey]

@HadrienGardeur Based on my research, most Japanese reading systems use the simple check of the class list of the HTML element containing a class that matches gaiji or gaiji-* (some check the filename but I wouldn't recommend that). This is because the EBPAJ conformance guide has the following:

Any publication that has the ebpaj conformance metadata in the OPF and is in Japanese (language) gets this special treatment of the gaiji images.
I know it's not "standard EPUB", but so many reflowable EPUBs in Japan have this behavior that if we want an significant improvement in the accuracy of the tool for Japanese EPUBs it should be added at some point.

[HadrienGardeur]

I see, but without an alternate UTF-8 character, I don't see how we could ignore these images.

[chocolatkey]

Yes, that's why I'm only suggesting it. I think for analysis of large batches of publications, it could help to have this as a flag to filter out the gaiji issues from the rest.