CA-LWNet: A Lightweight Network for Handwritten Dongba Character Recognition

Although significant progress has been made in handwritten Dongba character recognition (HDCR), current methods still fall short in promoting Dongba culture and developing lightweight models suitable for non-professional users. Research on lightweight HDCR remains in its early stages, and most existing models fail to balance recognition accuracy and parameter efficiency. To address these issues, we propose a novel lightweight network with an extremely low parameter count, CA-LWNet. It features a compact architecture that achieves top recognition accuracy among existing lightweight models. A four-branch feature extraction module (LCMR), incorporating a coordinate attention mechanism (CA), enhances key feature extraction through an original feature branch, a local attention (LA) branch, and a medium-range attention (MRA) branch. Additionally, we apply a data augmentation method that preserves the structural integrity of handwritten Dongba characters to improve model robustness. Despite its extremely low parameter count, CA-LWNet demonstrates outstanding performance across multiple evaluation metrics, achieving state-of-the-art results among current lightweight networks.