Dissimilarity Matrix: Bridging the Gap of Bangla OCR Error Correction Through a Novel Approach
Noor Masrur, Faias Satter, Sk. Md. Masudul Ahsan · 2023
Optical Character Recognition (OCR) holds vital importance across sectors, including the Bangla language, yet errors endure due to intricate handwritten characters and vast character classes. Effective error correction methods are thus pivotal in the OCR post-processing stage, with room for enhancement despite existing techniques. Levenshtein Distance with dictionary lookup and subsequent contextual analysis are widely employed methods. This study introduces the novel Dissimilarity Matrix for enhanced OCR error correction. This method serves as an intermediary layer between the Levenshtein distance and contextual analysis by storing the degree of dissimilarity between two characters' appearances in a matrix. Since deep learning techniques are fundamental to most Bangla OCR systems, the proposed system leverages the cause of the error to correct it, particularly in cases where two characters are similar in appearance. The dissimilarity matrix was computed using both machine-generated and empirical values. A comparative investigation was undertaken between two proposed variants of dissimilarity matrices and the existing methodologies. The proposed empirical method achieved the best average F1-score of 93.04 percent and surpassed all other techniques in terms of speed, with an average run time of 0.20 seconds per word for contextual analysis. Notably, the study only considered primary Bangla characters and numeric characters, and compound characters can be incorporated to improve performance. With its swift processing time, the empirical dissimilarity matrix holds excellent potential for real-world applications.