An insurance company is working on a system that can perform Optical Character Recognition (OCR) on documents from their archives. Unfortunately, their documents have been scanned at angles ranging from -5° to 5° from the horizontal. To increase OCR accuracy, they need a preprocessing step that determines the angle of a given page so this distortion can be corrected.
-
The input of learning algorithm will be the scanned document images and the labels which are the different possible scanning angles.
-
Labels have been provided, treat with the problem as regression problem to get accurate result as much as possible.
-
Here we are trying to solve this probelm using Classical approaches Hough Transform.
-
Using Hough Transform Algorithm to Detection and Correction Skewed Document.
-
Using OCR to see our result pytesseract model provided by Google.
-
Build web app using django.
-
Optional:- deploy Using heroku.
Data link:-
https://www.kaggle.com/datasets/sthabile/noisy-and-rotated-scanned-documents
Work with noisy and rotated document images which language Spanish that make OCR model perform poor. Get only words with accuarcy more than 60%.
Achieve with mean square error on 500 label data 0.09492877999439271 error loss without rounding angle with rounding get 0.054 error. Execution time less than: 300 ms ± 38.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each).
- Fast Hough Transform. Bezmaternykh, P. V., and Dmitry P. Nikolaev. "A document skew detection method using fast Hough transform." Twelfth International Conference on Machine Vision (ICMV 2019). Vol. 11433. SPIE, 2020.
- Skew Detection and Correction using Hough Transform. Muthukrishnan,Computer Vision,August 22, 2020.
- Skew correction in Documents using Deep learning. Vishnu Nandakumar.