In order to overcome the limitations of existing headdress segmentation methods in portrait Thangka images and the high cost of fully supervised semantic segmentation with pixel level annotation,we propose a weakly supervised semantic segmentation method with frame level annotation.Firstly,the proposed method uses Canny algorithm to obtain the rough edge of headdress.Secondly,the improved EDLines algorithm is used to extract the key points of headwear.Finally,we use Polygons processing to generate feature masks according to the characteristics of headwear.Experiments show that the mean intersection over union,mean intersection over union (mIoU) index of this method is 7.56% higher than semantic segmentation instance (SDI) and is 6.11% higher than weakly-supervised instance segmentation_bounding box prior (WSIS_BBTP).It is effective.