Character classification enhancement through hybrid feature fusion in challenging scripts systems

One of the most intriguing research problems is to achieve high accuracy in character recognition of degraded scripts, which is essential for applications such as document digitisation, language translation, and text-to-speech systems. We aim to recognise two degraded scripts of Devanagari and Urdu languages, which have unique difficulties, mainly due to the presence of broken and merged dots. Traditional character recognition techniques, including template matching and feature-based methods, have been widely used but need to be more efficient to handle the complexities of Urdu and Devanagari scripts. We propose classifying damaged scripts using zone-based power curve fitting and a pre-trained VGG19 model that trains on script degradation patterns. Using 6,250 printed examples with distortions from damaged Devanagari and Urdu manuscripts, we fine-tune the VGG19 model. It helps the proposed model understand these characters' intricate features and minimises overfitting. Our changes improve accuracy and strengthen the script damage detection system. Our results show that the VGG19 architecture works well across most feature extraction strategies, with accuracy scores ranging from 89.26% to 93.42%, while combining the power curve fitting methodology with VGG19 improves classification accuracy to 97.42%.