LMAFNet: lightweight multi-scale attention fusion network for handwritten Dongba character recognition

Dongba ancient documents are listed as “Memory of the World” by UNESCO, which shows its important role for global language research. However, current deep learning methods for handwritten Dongba character recognition (HDCR) prioritize accuracy over parameter efficiency, hindering practical deployment. To address this limitation, we introduce a novel Lightweight Multi-scale Attention Fusion Network (LMAFNet) for Dongba character recognition, which is designed to balance model capacity and recognition accuracy. First, a Multi-scale Attention Fusion Block (MAFB) employs three Scale Convolutions (ScaleConvs) with different kernel sizes to capture both local details and global information. Second, ScaleConv is a novel structure that compresses and expands feature channels, enhancing nonlinear feature extraction while reducing the number of parameters. Third, an Efficient Channel-Spatial Attention Mechanism (ECSAM) that utilizes efficient 1 D convolutions to emphasize discriminative features across channel and spatial dimensions. Experimental results demonstrate that LMAFNet outperforms state-of-the-art methods on public benchmark dataset.