AttGraph disentangling confusable ancient Chinese characters via component-correlation synergy

With three millennia of evolution history, ancient Chinese characters present unique challenges for intelligent recognition systems due to severe degradation and stylistic variations, causing notable intra-class variation and reduced inter-class separability. To address this, we propose a novel framework with two key components: a local content transformation (LCT) module enhancing salient regions through learnable spatial attention weights, and a local correlation reasoning (LCR) module employing graph convolutional networks (GCNs) to model patch-wise spatial-semantic dependencies via Transformer-derived attention maps. Experiments on the MACT benchmark show it achieves state-of-the-art performance with a 2.02% top-1 accuracy improvement compared to existing methods. The proposed architecture exhibits particular robustness in handling degraded samples and distinguishing visually similar character classes, showing substantial potential for applications in archeological documentation, digital paleography, and intelligent cultural heritage preservation systems. Code is available at https://github.com/wtq123-git/AttGraph .