JISE


  [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]


Journal of Information Science and Engineering, Vol. 39 No. 1, pp. 1-18


Weakly Supervised Semantic Segmentation for Headdress of Thangka Images

WENJIN HU1,2,+, JIAHAO MENG1,3, LI JIA1,3, FULIANG ZENG1,2 AND PANPAN XUE1,2
1Key Laboratory of China’s Ethnic Languages and Information Technology
Ministry of Education
Lanzhou, 730000 P.R. China
2School of Mathematics and Computer Science
3National Languages Information Technology, Boulder
Northwest Minzu University
Lanzhou, 730000 P.R. China
E-mail: hwjforwork@126.com+; mjhforwork@163.com; 743507179@qq.com; y181730416@stu.xbmu.edu.cn; y191730464@stu.xbmu.edu.cn

In order to overcome the limitations of the existing headdress segmentation methods for portraits Thangka images and the high cost of the pixel-level annotation is fully supervised semantic segmentation, we propose a weakly supervised semantic segmentation method with box-level annotations. Firstly, the proposed method uses the Canny algorithm to obtain the rough edge of the headdress. Then, the new method improves the EDLines algorithm to extract the key points of the headdress. Finally, we use Polygon’s processing to generate feature masks according to the characteristics of the headdress. Experiments show that in the segmentation of the headdress of Buddha in portraits Thangka pictures, the index mIoU of the proposed method has 7.56% higher than SDI and 6.11% higher than WSIS_BBTP of the segmentation result, which are two state-of-the-art methods.

Keywords: Thangka image, semantic segmentation, weakly supervised, the headdress of Buddha, the CEDLines-polygons

  Retrieve PDF document (JISE_202301_04.pdf)