[IQL HackDay Challenge 2021]

Problem Statement

Write a program/train an ML model to find rectangular blocks of text on a given image, the image may or may not have a background color and alpha opacity. The image will contain multiple news item in nonstandard layout. The text need not be processed and may even be ignored as not being text (can be just treated as area of interest). No OCR etc is required however if it will be a plus. This program will be used to identify text blocks in image newspapers and scanned documents.

Example Image 1 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-0.png

Desired output:

coordinates of rectangles in x,y,x1,y1 form where x,y is coordinate of top left corner and x1,y1 is coordinate of bottom right corner in the rectangle.

Each line represent one box on the same image, a value of +- 10 is still acceptable.

38, 1245, 788, 2411

788, 2039, 1565, 2412

790, 1599, 1372, 2039

1373, 1655, 1565, 2040

1525, 1881, 1526, 1883

1372, 1245, 1561, 1653

787, 1247, 1370, 1602

38, 984, 608, 1245

41, 296, 610, 999

611, 476, 811, 1245

814, 902, 1567, 1245

608, 308, 1586, 899

The final result is used in this manner:

http://epaper.roznamasahara.com//home.php?issueid=1043&edition=1&date=2021-02-03

More Images

Image 2 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-1.png

Image 3 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-2.png

Image 4 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-3.png

Image 5 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-4.png

Image 6 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-5.png

Image 7 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-6.png

Image 8 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-7.png

Image 9 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-8.png

Image 10 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-9.png

Image 11 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-10.png

Image 12 : https://epaper-pro-images.s3.ap-south-1.amazonaws.com/1043-11.png

We have csv records of over 2 years with 8/12 pages per day which may be used as training dataset.