Feature‐Driven Deep Architecture for High‐Precision Recognition of Complex Handwriting

ABSTRACT Handwritten character recognition (HCR) plays a crucial role in digitizing documents, especially in developing countries where much information remains in hard copy. While automation improves accessibility and preservation, challenges such as variations in handwriting style, size, and image quality persist. Traditional models struggle with these variations, and while deep learning models are accurate, they often lack efficiency and generalization. To address this, we propose an adaptive HCR framework combining advanced preprocessing, feature extraction, and bio‐inspired optimization. Our approach uses a novel HOG‐Saliency‐Structural Feature Map (HOG‐SSFMap) that enhances feature learning by integrating Histogram of Oriented Gradients, saliency, and structural features. For classification, we utilize a PSILO‐DNet model based on DenseNet121 optimized by Paniscus Social Interaction Learning Optimization (PSILO), inspired by primate group behavior. This model ensures stable and accurate convergence. Experiments on benchmark datasets show promising results, achieving 96.01% accuracy on Devanagari and 95% on PHDIndic_11, proving its scalability and effectiveness.