Recognition of damaged characters in inscriptions: an integrated SVM shallow machine learning and miniCPM-V 2.0 large model aided recognition method

Many ancient inscriptions have suffered damage due to natural and human factors, resulting in incomplete texts. These damaged inscriptions are crucial for understanding ancient culture and history. However, traditional character recognition techniques often struggle to identify the damaged script accurately. In this study, We propose an integrated method combining support vector machine (SVM) and the MiniCPM-V2.0 model to enhance the recognition accuracy and processing efficiency of damaged inscriptions in Li Shu during the Wei, Jin, and Southern and Northern Dynasties. We initially extract character images from damaged inscription rubbings and conduct deep learning and semantic analysis using MiniCPM-V2.0 to optimize the recognition results. Moreover, for complex images, we develop an underlying SVM algorithm for feature classification and learning. The constructed dataset contains several well-annotated images of Lishu inscriptions, including samples of various degrees of damage. Experiments on small-scale data show that MiniCPM-V2.0 can effectively identify inscriptions while improving identification efficiency. The results demonstrate that MiniCPM-V2.0 can provide a solid basis for identifying complete characters, demonstrating the effectiveness and potential of the selected technical routes.