InscriberNet: A Novel Framework for Chinese Character Recognition in Stone Inscriptions using CNNs and Swin Transformer

Abstract. Stone inscriptions pose distinct challenges for optical character recognition due to erosion, disruption, and fragmentation. We present InscriberNet, a deep learning network designed for the recognition of Chinese characters from degraded stone surfaces. It incorporates a CNN-based denoising module, a ResNet-based feature extractor, a Multi-Scale Attention Mechanism, and a Swin Transformer to capture both local and global data. InscriberNet, assessed using the Hwaeom Stone Sutra dataset, attains an accuracy of 86.5%, surpassing conventional models such as CRNN and CNN-based denoising OCR. The findings underscore its resilience, efficacy, and relevance for the digitization and preservation of culturally important historical documents.