OCR using traditional and DNN approach
Bhupendra Kumar, Sarvesh Tanwar, Shamik Tiwari · Mechatronics · 2025
The recent advancements in machine learning technologies have significantly impacted the performance, accuracy, and robustness of machine learning-based applications. This chapter explores the comparison between traditional and deep neural network (DNN)-based approaches for optical character recognition (OCR) systems, highlighting the significance of integrating modern techniques into OCR. The novelty of this study lies in its comprehensive evaluation of both methods on Bangla machine-printed document images. The primary objective is to analyze and compare the challenges and performance of OCR systems implemented using traditional image processing techniques versus those using DNNs. The traditional approach involves preprocessing and segmentation using image processing routines, representing symbol images with histogram of oriented gradient (HOG) features, and employing support vector machines (SVM) as the recognizer. In contrast, the advanced approach utilizes a DNN-based network to recognize word images. Both methods are rigorously trained and tested on a dataset of Bangla machine-printed documents This study provides a detailed performance analysis of two distinct OCR approaches, offering valuable insights into the effectiveness of DNN-based methods over traditional techniques. The comparative analysis addresses key challenges and evaluates accuracy of each method in terms of word recognition rate and character recognition rate. The results demonstrate that the DNN-based OCR system outperforms the traditional approach in terms of accuracy and robustness, establishing it as a more effective solution for OCR tasks in Bangla document images.