大连民族大学学报

为了解决石鼓文识别问题，提出了采用多尺度融合的识别网络，并使用自建的石鼓文数据集进行单字识别的实验与分析。该网络通过引入坐标注意力作为Efficient Net V2-S的注意力模块，提高了石鼓文的识别准确度。引入多尺度融合模块，提取不同尺度的石鼓文全局特征并融入识别网络，增强了模型的泛化能力，从而提升了对临摹版石鼓文的识别能力。实验结果表明：该网络对原石鼓文拓片的识别率达96.76%，对吴昌硕临摹版石鼓文的识别率为90.12%。证明该网络在石鼓文原版及临摹版中的识别能力优于其他网络，解决了石鼓文在实际应用场景中的识别问题。此外，自建的石鼓文数据集也为后续的石鼓文相关工作提供了数据支持。

To address the challenge of Stone Drum Inscription recognition, we propose a multiscale fusion recognition network and conduct single-character recognition experiments using a self-constructed Stone Drum Inscription dataset. The network incorporates coordinate attention as an attention module within EfficientNetV2-S, which enhances the accuracy of Stone Drum Inscription recognition. The introduction of a multi-scale fusion module enables the extraction of global features of Stone Drum Inscriptions at different scales, which are integrated into the recognition network. This approach improves the model's generalization capability, thereby enhancing its ability to recognize replica versions of Stone Drum Inscriptions. Experimental results demonstrate that the network achieves a recognition rate of 96.76% for the original Stone Drum Inscription rubbings and 90.12% for Wu Changshuo replica versions. The study proves that this network outperforms other networks in recognizing both the original and replica Stone Drum Inscriptions, effectively addressing the recognition challenges in real-world applications.Additionally, the self-constructed Stone Drum Inscription dataset provides valuable data support for future research related to Stone Drum Inscriptions.