Research On Semantic Segmentation Method For Headdress Of Portrait Thangka Images

Thangka is a kind of religious scroll painting,which occupies an important position in Tibetan culture.It is a window to explore Tibetan traditional culture and religious customs.The segmentation of important semantic objects in the image can help the public understand the content of the image.In this way,we can realize our high-level semantic cognition of visual content.Therefore,the semantic segmentation of Thangka images plays an important role in understanding the rich connotations of Thangka images and promoting the spread of Tibetan culture.In portrait Thangkas,the headdress worn by the main deity can show the duties and attributes of the Buddha,and the headdress can identify the main type of this Thangka.Therefore,this article takes the semantic segmentation of the main headdress in portrait Thangka images as the main research content.At the same time,in view of the scarcity of Thangka images and their strict composition and rich structural features,this thesis studies the semantic segmentation of headdress.The main research is as follows:(1)The construction of the data set.In recent years,the research of Thangka images has gradually become a hot spot,but there is no data set for the research of semantic segmentation of Thangka images.This thesis collects relatively high-quality portrait Thangka images from the Thangka images obtained from school and network resources,the wizard annotation assistant and the self-built Thangka image annotation system are respectively used to mark the image at pixel-level and box-level.On this basis,a box-level headdress annotation data set with 883 images and a pixel-level Thangka headdress annotation data set with 552 images have been constructed.(2)In view of the problem that pixel-level annotation consumes a lot of manpower in fully-supervised semantic segmentation,we propose a weakly-supervised semantic segmentation method for headdress based on border-level annotation.The method first uses the Canny algorithm to obtain the rough edges of the headdress in the marked area.Then,it uses the improved EDLines algorithm to extract the key points of the headdress.Finally,Polygons Processing is designed according to the characteristics of the headdress to generate a more refined marking area of the headdress.Experimental results show that the method is superior to the two SOTA-level frame-level weakly supervised semantic segmentation models of SDI and WSIS?BBPT in terms of the segmentation of Thangka image headdress,and m Io U improves by 7.56%and 6.11%.(3)Considering that thangka images obtained by digital acquisition are characterized by high noise,low image quality and high complexity.In the task of Thangka image segmentation,traditional image segmentation methods have the problems of insufficient accuracy and adaptability.First,the model uses FDog to enhance the edge of the Thangka image.Secondly,the feature map is extracted through a feature extraction network composed of Res Net and Bi FPN.Third,use the Half-RPN module designed according to the characteristics of Thangka image composition to obtain the recommendation frame(ROI).Finally,we use Mask R-CNN Head to complete segmentation and category prediction.Experimental results show that the performance of this model is better than mainstream models such as Mask R-CNN,Deep Lab V3~+,and Cascade Mask R-CNN,the m Io U is improved by19.89%compared with DeepLab V3~+.