COMBINING CLASSICAL AND DEEP LEARNING METHODS FOR EFFECTIVE TEXT IMAGE MANIPULATION

Abstract

Abstract—Text recognition and segmentation on a wide variety of images is crucial for applications such as document analysis, autonomous exploration, multimedia content encoding, etc. However, text recognition accuracy is typically hampered by noise, low resolution, and complicated backdrops. In this research, we describe a novel strategy that mixes classical image processing approaches with deep learning techniques, simultaneously exploiting their strengths to construct a text segmentation model. Classical approaches utilized here are adaptive thresholding, morphological operations, and edge detection, which work well for preprocessing and enhancing text clarity. Simultaneously, deep learning architectures (such as U-Net and Mask R-CNN) extract features, segment text, and recognize bounding boxes for bounds of text. So, using classical image processing techniques such as morphological transformations of the edges, transformations in both standard and gradient shape, etc., we can improve the quality of the input image to use the deep learning models to segment the input image precisely and also localize the regions of text with high accuracy. Main contributions include establishing a comprehensive framework for recognizing text causes. We produce fantastic results, obtaining a max precision of 95.4% combined with a recall of 92.3%, an F1 score of 93.8%, and a tremendous score of 72.3%, indicating its supremacy over the state-of-the-art approaches. These outputs establish the efficiency of the combined model. We also draw key conclusions on the complimentary nature of both approaches and how they might inform each other and illustrate the work we undertook to adapt classical text segmentation algorithms to all major circumstances for picture data ranging from text-based to feature/texture-based ones. The suggested system has practical applications that can extend not only to the optimization of models but also to evaluating performance on a bigger dataset in real time.