OCR0052: Automating Glyph Extraction from OPF eText

[10kalden]

Description:

• We are extracting glyphs by scanning the OPF text and determining on which line the character occurs
• We use pagination to find the image on which this character occurs, we use this data to load on Prodigy and use annotators to crop the character. This work is done here in this https://github.com/orgs/OpenPecha/projects/56/views/7?pane=issue&itemId=67820570
• To scale this work, we need to find a way to automate this process without the need for annotators

Implementation:

The approach for this has already been tried: indexing the text and cropping the character by expanding its bounding poly. However, this was not optimum, and the results were not good.

• The other approach is to find which quadrant the character occurs and crop that quadrant.

Estimation: Initial Estimation: 4 days start date: 23-08-24 end date: 28-08-24

Updated Estimation: 3 days start date: 03-09-24 end date: 05-09-24

Sub-task:

• [x] Divide the text and the images into segments and match them.
• [x] Write a script segment the image and crop the required segment
• [x] Scan the entire OPF to get character segment info
• [x] Use the layout analysis model the crop the text area

Completion Criteria: To extract all the glyphs without the need for manual annotation

[10kalden]

For character ཅ, in the 4th volume of Kangyur

img.1 Original Image

img.2 the character found in 2nd quadrant

img.3 the character found in 4th quadrant

[10kalden]

• This method of dividing the Pecha image into quadrants and cropping the area yields a large area of the Pecha for a single character.
• To solve this, I will further divide the area of text quadrant into smaller area and do the same to the image also, and try it that way

[10kalden]

• Divided the text into top and bottom half, further divided it into 8 segments.
• This yields better results than quadrants but needs some exploration and solutions to get the exact glyph

img.1 original image

img.2 cropped image for ཅ_1

img.3 cropped image for ཅ_2

img.3 cropped image for ཅ_3

[10kalden]

• To find char ཅ using this method in a Pecha image with 5 ཅ, only around 2 or 3 cha can be obtained due to

Issue.1 The border around the images is causing problems with correct image segmentation.

Possible Solution

1. Tesseract to detect text area
2. openCV
3. Cropping the image from with a fixed value

[kaldan007]

test with our layout analysis model

[10kalden]

• Used the layout analysis model to crop the text area

• ཅ variant in 1 one page

img. ཅ_1

img.ཅ_2

img.ཅ_3

img.ཅ_4