Traditional Pattern Segmentation Algorithm Based on Semantic Prediction and Region Learning*

Traditional cultural images have large differences in style and details in their each subdivision field, which makes the existing segmentation algorithms based on deep learning models difficult to achieve ideal segmentation results. Based on the similarity of structural information of traditional cultural images, we propose a traditional pattern semantic segmentation algorithm based on semantic prediction and region learning, which improves the segmentation accuracy of traditional cultural image that insufficient pixel-level labeled training data. Firstly, a primary classification model is trained by using the category labels annotated at the image level, and the classification network is used to generate the category activation map to realize the classification and location of the patterns in the image. Then, the segmentation pseudo-label was used as the initial seed region, and the seed region was expanded by region growing of the seed point to obtain the pixel-level segmentation pseudo-label annotation. Finally, the pixel-level pseudo-label labeling information was used to obtain the final pattern segmentation results through the network model optimized by the dynamic supervised training strategy. The quantitative and qualitative experimental results show that, compared with the state-of-the-art the same category algorithms, our algorithm achieves good experimental results on different types of traditional cultural image patterns, the mIoU reaches 49.8%.