EffiHCR-Net: A Lightweight Hybrid CNN–Attention Architecture for Accurate, Efficient, and Style-Robust Handwritten Character Recognition

Abstract

Handwritten character recognition remains challenging due to variability in handwriting style, visually similar characters, and the difficulty of segmenting connected or cursive strokes. This study presents EffiHCR-Net, a compact convolutional neural network combining depthwise-separable convolutions, a residual connection, and Squeeze-and-Excitation channel attention, trained and evaluated on the EMNIST Balanced dataset (47 classes; 101,520 training, 11,280 validations, and 18,800 test images). EffiHCR-Net achieved 88.45% test accuracy (macro F1 = 0.8838) using 84,519 parameters, an 80.20% reduction relative to a conventional baseline CNN (84.93% accuracy; 426,799 parameters), though at a measured 16.50% higher CPU inference latency. An ablation study isolating the attention block showed a +1.40 percentage-point contribution. On a photographed real-world handwriting sample, the character-segmentation pipeline produced degraded output due to connected cursive strokes, whereas a pretrained Transformer-based recognizer (TrOCR) correctly transcribed the same sample.