enhancement: bring back embedded images in PDF's

[christinestraub]

While the fix for the PR that addressed #176 was a big move forward in merging OCR bboxes+text with layout-model bboxes+text, the issue remains that we are still dropping the "Image" element category that was lost in https://github.com/Unstructured-IO/unstructured/pull/1142/files / de19ace. Actually, in the comment there

        # Skip extracted images for this purpose, we don't have the text from them and they
        # don't provide good text bounding boxes.

Users may want the Image bbox info anyway even if there is no text. One thing we are doing correctly now that shouldn't change is full-page embedded Image PDF's. (e.g. the IRS pdf in ingest tests).

Extracted bboxes:

Final bboxes:

[badGarnet]

would a reasonable outcome of this request being:

• image elements are preserved in the list of output elements here

  unstructured-inference/unstructured_inference/inference/layoutelement.py

  Line 131 in de19ace

  extracted_elements_to_add.append(extracted_region)

  without having it go through the merging logic?

[cragwolfe]

yes, as long as the image itself is not the entire page (or worst case, there is an arg to specify the behaviour).

…

On Tue, Sep 5, 2023 at 7:35 AM Yao You ***@***.***> wrote: would a reasonable outcome of this request being: - image elements are preserved in the list of output elements here https://github.com/Unstructured-IO/unstructured-inference/blob/de19ace101632ed5a8b8e0fed62cef0795fd55ac/unstructured_inference/inference/layoutelement.py#L131 without having it go through the merging logic? — Reply to this email directly, view it on GitHub <#194 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AG2BGJ5SSJLEOMQ77U7QTGLXY42C3ANCNFSM6AAAAAA4K3NNSQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.*** com>