SCENE TEXT IMAGE SUPER-RESOLUTION USING DEEP LEARNING

Abstract

Scene Text Image Super-Resolution (STISR) is a vital sub-task in document analysis and optical character recognition (OCR), particularly in real-world applications such as license plate recognition, translation, assistive technologies, and augmented reality. In unconstrained environments, text captured in natural scenes often su!ers from motion blur, low resolution, defocus, and other degradations, making it di"cult for both humans and machines to interpret the content accurately. Traditional superresolution techniques, primarily designed for natural images, often fail to preserve the fine-grained semantic and structural features of text, such as strokes, curves, and character boundaries. This thesis addresses these challenges by proposing a suite of di!usion-based generative models that incorporate structural, semantic, and linguistic priors to restore high-quality, readable text images. The first contribution introduces the Di!usion-Conditioned Di!usion Model (DCDM), which integrates low-resolution visual input with character-level semantic embeddings derived from a latent text di!usion module. The second contribution presents the Skeleton-Aware Di!usion Model (SADM), where refined skeleton maps act as structural guides during the di!usion process to improve character-level fidelity. The third and final contribution explores on the stroke level by the Stroke-Conditioned Di!usion Model (SCDM), focusing on low-resource Indic scripts such as Hindi, Marathi, and Gujarati. SCDM leverages a novel stroke decomposition strategy to generalize across linguistically similar scripts through structural supervision.