Research And Application Of Automatic Detection Of Thangka Elements Based On Deep Learning

As the main form of Regong Art,thangka is an important art carrier of Tibetan Buddhism.It includes history,religion,Tibetan medicine and other aspects.The study of Thangka can better protect,inherit,popularize and carry forward the Qinghai Tibet culture and understand the historical features.For images such as Thangka,which are all inclusive,colorful,complex and mysterious,relying on the traditional manual method to recognize the target features of Thangka needs to reserve a lot of a priori knowledge of Thangka,and it will consume a lot of time and energy.Combine the learning of Tangka network to realize the automatic detection of a large number of unmarked image elements,and then combine the learning technology of Tangka network to realize the automatic detection of a large number of Tangka images.It has important cultural significance and academic value for the popularization,protection,inheritance and development of Thangka art.Based on the automatic detection of Regong Thangka elements and improving the detection accuracy,this paper puts forward a target detection method more suitable for Thangka images.The main work of this paper is as follows:（1）At present,there is no authoritative and open standardized Regong Thangka data set for in-depth learning in the field of image target detection.Therefore,this paper constructs Thangka image data set and label data set applied to the field of image target detection.Aiming at the magic instrument elements with inclined position and large difference in length width ratio in Thangka images,this paper also creatively constructs the Thangka rotation target data set with rotation detection frame,which fills the gap of Thangka data set in the field of target detection.After screening,data enhancement and other processing,there are about 2000 standardized Thangka images in the data set,which can meet the research requirements in the field of in-depth learning.（2）Under the condition of ensuring the accuracy of target detection in Thangka image,in order to make the model lighter,faster detection speed and more suitable for Thangka image,this paper improves the yolov5 model,introduces the ghost model,replaces part of the original bottleneck convolution module with the ghost bottleneck module,and reconstructs the ghost bottleneck CSP module combined with the C3 module in the original model,And replace the ordinary convolution with ghost convolution.The experimental results show that the improved network structure has greatly improved the detection accuracy of Thangka targets,greatly reduced the amount of parameters of the model,and significantly improved the image detection speed.（3）In view of the situation that the detection accuracy of inclined elements with large difference in length width ratio in Thangka image is not high or even missed,this paper adds a rotation detection frame（rbox）,and accurately detects Thangka elements with horizontal detection frame and rotation detection frame.CSL（circular smooth label）technology is introduced based on YOLOv5-Ghost Module model,Change the angle problem from regression problem to classification problem.Experiments show that after adding the rotating target detection frame and introducing CSL technology,the objects that are not easy to be detected in Thangka can also be detected.The detection effect of Thangka rotating objects has been significantly improved and the accuracy has been improved to a certain extent.（4）Based on the improved model,this paper designs and implements the automatic detection system of Thangka elements.The function of this system is convenient to operate and can meet the requirements of automatic detection of Thangka elements.It is conducive to the popularization and development of Regong Thangka art.