The exponential growth of unstructured and heterogeneous documents in fields like healthcare, finance, legal systems, and education has further exacerbated the demand for reliable automated text extraction systems. Traditional OCR tools have difficulty processing noisy, distorted, or degraded scanned documents, in which artifacts, such as stains, blur, irregularities in handwriting, shadows, and different layouts, make recognition accuracy extremely difficult. To solve these shortcomings, this study proposes a novel integrated framework based on generative adversarial networks (GANs) and YOLOv8-based text region detection and post-OCR correction mechanisms. In the proposed system, GANs are used to generate various and realistic document variations, including multiple fonts, languages, degradations, and document structures, which reduce the problem of dataset scarcity and enhance generalization in low-resource or privacy-restricted scenarios. YOLOv8 is used to accomplish the accurate localization of text and stamp regions, even in a cluttered or occluded background, thus improving the quality of segmentation before OCR processing. Furthermore, semantic-aware post-OCR correction refines the extracted text by leveraging contextual modeling to reduce common OCR errors, such as substitutions, deletions, and character-level noise. This integrated pipeline enhances the textual output to be substantially more faithful to the original content while minimizing the propagation of errors into downstream applications. Such improvements are particularly important for systems based on natural language processing (NLP) and retrieval-augmented generation (RAG), which are highly sensitive to input quality. Empirical evaluations demonstrate notable performance gains over baseline OCR models, especially under degraded document conditions, thereby validating the effectiveness of GAN-enhanced preprocessing combined with YOLOv8-based detection. The proposed GAN-YOLOv8-OCR system provides a robust, scalable, and domain-agnostic system to digitalize multilingual historical, and other complex scanned documents, and, in the process, enhances intelligent information retrieval and decision support systems.
Sreevidya et al. (Sun,) studied this question.