Research on Semantic Segmentation of Thangka Images

Thangka is one of the important artistic works in Chinese culture. In order to better promote the understanding and research of digital thangka images, this article explores the semantic segmentation of the headdress of the main statue of thangka. In this article, we use FCN as the benchmark model, and choose PSPNet, Deeplab V3+, Mask R-CNN, DANet, Cascade Mask R-CNN and other five mainstream semantic segmentation models based on deep learning. Then carry out experimental research separately. We explore a more suitable deep subdivision research model for Thangka image semantics. The construction of Thangka image semantic segmentation model provides experimental and theoretical basis for the next step. In this paper, using the same parameters and not using the pre-training model, the conventional training experiments and comparative training experiments under different backbone networks were carried out. Experiments show that using FCN as the baseline model, cascaded Mask R-CNN has the best segmentation effect in headwear segmentation. Its mIoU is 62.59%. In different backbone networks, PSPNet segmentation effect is significantly different.