Multi-resolution feature integration framework with CNN for Bengali handwriting quality assessment

Abstract

This paper proposes a multi-resolution CNN framework for automated assessment of Bengali handwriting quality. The model captures stroke-level details, curvature and junction patterns, and overall character structure through three parallel convolutional streams. Tested on Bengali handwritten character images, the framework achieved 96% accuracy and 95.8% precision, outperforming several standard CNN models while using only 8.2 million parameters. The study demonstrates its potential for automated handwriting evaluation and educational feedback systems.