Object Detection of Thangka Image Based on Improved YOLOX

Thangka is a unique painting art of Tibet and a world intangible cultural heritage. The background of Thangka images is complex, the picture content is rich, and there are many objects. We first constructed a Thangka Multi-Object Image Dataset by annotating the locations of religious tools, seats, and headdresses in 2079 images of Thangkas. Moreover, based on this, we conducted object detection research for the religious tools, seats, and headdresses in the Thangka, introduced Dynamic Head and CIoU Loss into YOLOX, and proposed an improved model TK- YOLO. It is experimentally demonstrated that the proposed method obtains an average accuracy of 54% on the Thangka Multi-Object Image Dataset, improves the performance of the baseline model by 1.9%, and has a significant advantage over other classical object detection models with TK-YOLOX.