Extraction of Buddha Elements from Thangka Images Based on Improved DeepLabV3+

Nowadays, Thangka as China's intangible cultural heritage has attracted more and more attention. Semantic segmentation of Buddha elements in Thangka images can let the public understand Thangka culture. Using the results of semantic segmentation, the artistic elements in the image can be extracted and recreated as material. This paper constructs the first semantic segmentation dataset of Buddha elements based on Thangka image, and proposes an improved DeepLabV3+ model based on DeepLabV3+ model. Firstly, the lightweight MobileNetV2 is used for feature extraction, and then the asymmetric convolution module is used to replace the dilated convolution in the ASPP block to obtain richer features, and the segmentation accuracy is guaranteed without increasing the calculation time. Finally, the segmentation results are used to extract the Buddha elements in the Thangka image. Experiments show that the improved DeepLabV3+ model proposed in this paper has higher accuracy and less calculation than the original DeepLabV3+ model on the Thangka dataset.