Advancing Handwritten Digit Recognition: A Hybrid Deep Learning Framework with Attention Mechanisms

Abstract

Handwritten digit recognition is a pivotal challenge in computational vision,demanding robust models to interpret diverse handwriting styles. This study proposes a novel hybrid deep learning framework integrating convolutional neural networks (CNNs) with attention mechanisms to enhance classification precision andcontextual adaptability. The methodology encompasses preprocessing for featureenhancement, a tailored architecture combining CNNs, logistic regression, and random forests, and advanced statistical validation. The framework achieves a theoretical accuracy of 99.2%, surpassing conventional models. Statistical analyses,including Bayesian inference, ANOVA, and permutation importance, substantiatemodel robustness and feature significance. By replacing traditional feature importance plots with probabilistic coherence metrics, the study offers interpretableinsights into model performance. A conceptual prototype demonstrates scalability for applications in automated banking, document processing, and educationalassessment. This research advances machine learning paradigms, contributing toscalable, high-precision digit recognition systems.Keywords: Handwritten Digit Recognition, Convolutional Neural Networks,Attention Mechanisms, Deep Learning, Statistical Inference, Computer Vision