<p>Rock art is recognized globally as significant cultural heritage. Symbols in rock art capture scenes of daily life from ancient societies, revealing the cultural context of past civilizations and holding significant research value. In the research of rock art symbols, it is necessary to accurately and efficiently segment the symbols in the images to ensure subsequent research on rock art symbols and the construction of symbol system databases. Although existing methods for rock art symbol segmentation can effectively extract symbols from 2D images, they are often time-consuming, labor-intensive, and have low segmentation accuracy of the model. To address these challenges, this study proposes a rock art symbol segmentation method based on an improved YOLOv7-Seg model, which incorporates SE (Squeeze-and-Excitation Networks) and ODConv (Omni-Dimensional Dynamic Convolution) to enhance the model’s focus on rock art symbol features, enabling efficient and accurate segmentation in complex backgrounds. This model facilitates the recognition and segmentation of human and animal symbols in images. The study employs Cangyuan rock art as a case study, validating the model’s accuracy through ablation experiments and comparative analysis. The model achieves an overall AP score of 0.961, with specific AP score of 0.973 for animal segmentation and 0.948 for human segmentation. The results demonstrate that the improved model effectively segments rock art symbols in complex environments, achieving high-precision automated segmentation of rock art symbols. This research lays the foundation for the subsequent unified management of rock art symbols as well as the study of rock art protection and heritage preservation.</p>