ShuiAttNet: Fully convolutional attention network for Shuishu character recognition

Abstract

Shuishu is one of the most representative hieroglyphs and precious cultural heritage in China, currently facing the risk of extinction. Preserving this endangered script requires innovative approaches to accurately recognize its characters. However, existing methods face significant challenges, including the inability to handle the broad diversity of Shuishu characters and the complexities of authentic ancient manuscripts. To address these issues, we present a comprehensive study that combines dataset construction and advanced deep learning methods. First, we establish the largest and most diverse Shuishu single-character dataset named S842 to date, addressing the critical lack of publicly available resources for Shuishu. Then we propose a novel Fully Convolutional Attention Network named ShuiAttNet, which is specifically designed for Shuishu character recognition. ShuiAttNet introduces two key innovations: the Attentional MBConv (AMC) block and the Fully Convolutional Attention (FCA) block. The AMC block utilizes a novel feature fusion mechanism to capture fine-grained local details while reducing feature redundancy caused by the low-rank characteristics of Shuishu characters. Meanwhile, the FCA block employs Depthwise Separable Dilated Convolution to establish long-range dependencies while preserving the two-dimensional spatial structure of the images. These components enable ShuiAttNet to achieve superior performance with significantly fewer parameters compared to existing methods. Extensive experiments validate the effectiveness and superiority of ShuiAttNet in both quantitative and qualitative assessments. Experimental results show that our proposed model achieves a Top-1 Acc of 97.04%, outperforming other state-of-the-art methods.