DeepArabicDoc: A Hybrid CNN-LSTM Framework for Multi-Scale Enhancement and Semantic Analysis of Historical Arabic Manuscripts

This paper introduces the novel structure for processing historical Arabic documents in integrating multi scale enhancement skills with semantic analysis. The proposed approach utilizes the combination of modern image processing techniques and deep-learning models including LSTM and CNN architectures to these challenges of degraded historical manuscripts. The framework employs the two-stage methodology: multi scale document development for improving the text readability and feature extraction optimization to boost recognition accuracy. Experimental validation on the VML HD dataset demonstrates a significant 30% improvement in feature extraction accuracy compared to baseline methods with the development of module successfully handling 76.9% of heavily degraded documents. The LSTM model achieved a test accuracy of 91.09% and a loss of 0.2926 showcasing its effectiveness in capturing semantic dependencies. With an average processing time of 1.5 seconds per page and this framework is efficient and scalable highly suitable for large scale digitization and analysis of historical Arabic texts. These results highlight the potential of the proposed system to increase the preservation and accessibility of cultural heritage.