OCR0008: Glyph creation by annotations (5)

[10kalden]

Description:

• After generating fonts using glyphs previously annotated by annotators, some Tibetan characters and ligatures haven't been annotated yet, requiring additional annotations.
• The pipeline for glyph extraction is already in place, but it can be optimized to reduce wait times for annotators.
• The essential list of Tibetan characters and Ligature is already present and could be used for reference.

The process involves several steps:

• Download the images from the publisher's work and use OCR to extract the textual content.
• Parse the OCR output using the glyph extraction pipeline to crop or isolate the specific glyphs relevant to the font creation.
• Present the cropped glyphs to annotators via GitHub to select the ten best glyphs for each desired Tibetan character and ligature.
• Upload the glyphs to the S3 bucket and then create a JSONL file to structure the data for loading into Prodigy.
• Load the JSONL file in Prodigy to annotate the glyphs.

Completion Criteria:

• A well-annotated glyph for all the desired Tibetan characters and ligature, ready to be integrated into the font creation pipeline.
• Optimised pipeline for glyph extraction to reduce wait time for annotators.

Implementation plan:

Sub-Task:

• [x] Write a script to find all the missing glyphs from the essential list
• [x] Familirise with the script
• [x] Write test module for downloading images from mock S3
• [x] Refactor the script and find the missing glyphs from OCR output JSON
• [x] Download the images from the s3
• [x] Upload the glyphs to GitHub to images selection by the annotators
• [x] Upload to s3 bucket
• [x] Create JSONL file to structure the data for loading in prodigy

[10kalden]

• Before proceeding with the work, we need to determine whether we are taking random images from multiple books like we are doing now or whether we are taking images from a single book from a publisher.
• Creating a font by taking random images from multiple books generates an inconsistent font.

[10kalden]

• found 91 missing glyphs from work id W3CN20612 for derge font
• cropped around 100 images for each glyphs
• need to find the remaining from the other works from derge
• currently finding the glyphs from tengyur pecing

[10kalden]

• cropped the glyphs from alread present ocr output and images from 5 publishers work
• need more books to get as much glyphs as required.

[10kalden]

• Need to use ec2 to download the images from the s3, the script is already there, setting up the env for the script to run

[10kalden]

• need some extra work from Derge, the images are downloaded, currently running OCR on the 5 more books. after that, the pipeline to crop the glyphs and uploading to Git Hub will be run on the server.

• repo created for glyphs https://github.com/MonlamAI/F20004

[10kalden]

• we are going to do all the glyphs process, annotation and font creation from Derge Publications first, if it's successful, then we will process all the images from other publications.