Text spotting and lexicon based recognition of cursive handwriting in document images

Abstract

Computer vision is a dynamic field of artificial intelligence focused on enabling computers to interpret and understand visual data. Its applications span various domains, including image recognition, object tracking, newlineautonomous vehicles, and medical image analysis, reshaping how machines perceive the visual world. Handwritten Text Recognition (HTR) complements this by converting handwritten text into digital format, newlineleveraging machine learning and computer vision to transcribe accurately. Its versatility extends to digitizing historical documents, automating forms, and enhancing accessibility. Offline HTR, crucial in this study, involves analyzing static images of handwritten text after its creation. Its significance lies in its wide applicability across domains like digitizing historical documents, processing medical records, and automating paperwork processes. However, HTR poses challenges due to handwriting variations and image noise, which traditional methods struggle to address. Deep learning, a subset of machine learning, offers a solution by leveraging Neural Networks (NNs) to discern complex patterns in noisy data. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) excel in HTR tasks, with CNNs capturing spatial features and newlineRNNs modeling sequential relationships between characters. The Connectionist Temporal Classification (CTC) loss function is pivotal, facilitating the recognition of characters even in noisy or misaligned text. newlineThis thesis delves into various deep learning models, proposing novel architectures and methodologies tailored for HTR. It explores word spotting techniques, advances in line-level and page-level recognition, and integrates innovative approaches like the LexiconNet and gated LexiconNet. The culmination is a comprehensive discussion of research outcomes, paving the way for enhanced HTR applications across diverse fields.