BN-DRISHTI: Bangla Document Recognition through Instance-level Segmentation of Handwritten Text Images
This addresses the challenge of document recognition for Bangla, a widely spoken language, by providing a high-performance segmentation method and an extended dataset, though it is incremental as it builds on existing techniques like YOLO.
The paper tackles the problem of line and word segmentation in handwritten Bangla documents by introducing BN-DRISHTI, a method combining YOLO with transformations, achieving F-scores of 99.97% for line and 98% for word segmentation on their dataset and outperforming on external datasets.
Handwriting recognition remains challenging for some of the most spoken languages, like Bangla, due to the complexity of line and word segmentation brought by the curvilinear nature of writing and lack of quality datasets. This paper solves the segmentation problem by introducing a state-of-the-art method (BN-DRISHTI) that combines a deep learning-based object detection framework (YOLO) with Hough and Affine transformation for skew correction. However, training deep learning models requires a massive amount of data. Thus, we also present an extended version of the BN-HTRd dataset comprising 786 full-page handwritten Bangla document images, line and word-level annotation for segmentation, and corresponding ground truths for word recognition. Evaluation on the test portion of our dataset resulted in an F-score of 99.97% for line and 98% for word segmentation. For comparative analysis, we used three external Bangla handwritten datasets, namely BanglaWriting, WBSUBNdb_text, and ICDAR 2013, where our system outperformed by a significant margin, further justifying the performance of our approach on completely unseen samples.