Research On Thangka Image Object Detection Algorithm Based On Deep Learning

As a cultural relic treasure of the Chinese nation,Thangka images are related to religion,history,folk customs and many other aspects.Object detection on the headgear and seating of Thangka images can not only effectively help people understand the rich connotations contained in the images,but also provide a foundation for promoting a cultural power.Because Thangka images have the characteristics of complex background information,complex and diverse appearance of detection targets,and low image quality,higher requirements are put forward for target detection in Thangka images.The target detection algorithm based on deep learning has achieved fruitful results.However,due to the particularity of Thangka images,there is still room for improvement in detection accuracy and speed in the detection of headgear and seating.In view of the above problems,this paper selects the headgear and the seat in the Thangka image for detection research.The main research contents and innovations are as follows:(1)Build a Thangka image object detection dataset.Collect 883 Thangka images from the school library and the Academy of Fine Arts,and process them into a unified format.Secondly,use the wizard labeling assistant to label the headgear and seating of the thangka images in a rectangular frame.Finally,analyze the labeling data set and find that each sample Extremely unbalanced between categories.The Thangka image is enhanced by flipping,Gaussian noise and random occlusion,so as to improve the generalization ability of the detection algorithm in the Thangka image target detection.(2)In view of the characteristics of Thangka images with complex background information,complex and diverse appearance of detection targets,and low image quality,it is difficult for conventional detectors to effectively detect targets and there is a problem of missed detection.In this paper,based on the YOLOv5 detection algorithm,a target detection algorithm with dual optimization of hierarchical visual Transformer and attention mechanism is proposed.Firstly,the input image is sliced by the Focus module,and the sliced image features are extracted by the Swin Transformer.Secondly,the feature information of different layers is fused with the convolution kernel to obtain the multi-scale feature map.Finally,the self-attention mechanism is used to fuse the Neck module features are processed.The experimental results show that the detection accuracy of the improved YOLOv5 algorithm is 5.12%higher than that of the original YOLOv5 algorithm,and its detection accuracy is also better than other mainstream target detection algorithms.(3)Considering that the fixed anchor in the YOLOv5 algorithm damages the universality of the algorithm,and a large number of candidate frame generation also affects the performance of the detection algorithm.This paper further combines the characteristics of Thangka images,and takes Anchor free YOLOX as the baseline,and proposes an improved YOLOX Thangka image object detection algorithm.Firstly,the feature map extracted by the Darknet53 feature extraction network is fused using the coordinate attention mechanism,which strengthens the long-range dependency information and precise location information.Secondly,a new feature fusion method,Trans FPN,is proposed,which can perform better in feature fusion.It enriches the semantic information of global features and context features,and improves the accuracy of target detection.The experimental results show that the improved YOLOX algorithm has better detection accuracy,and the detection accuracy on Thangka images is improved by 6.4%.