Bangla handwritten word recognition using YOLO V5

Abstract

This research paper presents an innovative solution for offline handwritten word recognition in Bengali, a prominent Indic language. The complexities of this script, particularly in cursive writing, often lead to overlapping characters and segmentation challenges. Conventional methodologies, reliant on individual character recognition and aggregation, are error-prone. To overcome these limitations, we propose a novel method treating the entire document as a coherent entity and utilizing the efficient you only look once (YOLO) model for word extraction. In our approach, we view individual words as distinct objects and employ the YOLO model for supervised learning, transforming object detection into a regression problematic to predict spatially detached bounding boxes and class possibilities. Rigorous training results in outstanding performance, with remarkable box_loss of 0.014, obj_loss of 0.14, and class_loss of 0.009. Furthermore, the achieved mAP_0.5 score of 0.95 and map_0.5:0.95 score of 0.97 demonstrates the model’s exceptional accuracy in detecting and recognizing handwritten words. To evaluate our method comprehensively, we introduce the Omor-Ekush dataset, a meticulously curated collection of 21,300 handwritten words from 150 participants, featuring 141 words per document. Our pioneering YOLO-based approach, combined with the curated Omor-Ekush dataset, represents a significant advancement in handwritten word recognition in Bengali.