Improving Character Recognition in Bangla Handwritten Words: A Two-Stage Single Shot Detector Approach

Optical Character Recognition (OCR) for handwrit-ten documents is an important task for various applications such as digital archiving, document retrieval, and text-to-speech systems. Handwritten Bangla script, with its complex characters, poses a significant challenge for OCR systems. A two-stage approach using Single Shot Detector (SSD) with transfer learning is proposed to detect and recognize handwritten Bangla words from an image. In the first stage, a pre-trained model is fine-tuned on a custom dataset of 53 Bangla character labels. In the second stage, The character detection model is used as a backbone and fine-tuned on a custom dataset of Bangla handwritten word images, which also has the same 53 labels. Experimental results show that the two-stage approach achieves an F1 score of 82.86% on test set, outperforming traditional image processing techniques. This proposed approach can contribute to the development of more efficient and accurate OCR systems for Bangla documents, with potential applications in digital archiving, document retrieval, and text-to-speech systems.