Hybrid Architecture for Image Text Recognition: Applications in Industry 5.0 and Consumer Electronics Within 6G Networks

Scene text recognition technology has emerged as a pivotal conduit between the physical world and digital information, particularly within industrial settings where the variability of environments renders the recognition of low-quality scene text exceedingly challenging. This challenge is compounded by issues such as image blur, character similarity, occlusions and low resolution. This paper introduces a hybrid architecture that integrates encoder-decoder and diffusion models. Utilizing a dynamic adaptive method for image fragment embedding, this architecture preserves the integrity of character-level semantics. Additionally, the developed local attention mechanism, which calculates attention based on feature proximity, significantly enhances the recognition capabilities for visually similar characters. By incorporating a supervised diffusion process and employing varied denoising strategies across different temporal steps, the model adeptly handles visual information in degraded scenes without reliance on image preprocessing or data augmentation. Validation on the TextZoom dataset, across hard, medium, and easy blur levels, shows that the proposed method’s accuracy surpasses other advanced methods by 10.8%, 8.5%, and 5.7%, respectively. In experiments on an industrial metal parts marking dataset, it exceeds by 6.5%. Code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/xia0shutng/Scene-Text-Recognition</uri>.