Study on the Headgear and Seat of the Thangka Image based on the Improved YOLOv4 Algorithm

Abstract

Aiming at the problems of complex background and low accuracy in Thangka image detection tasks using deep learning algorithms, an improved object detection algorithm based on YOLOv4 is proposed. Considering that the Mish function used by YOLOv4 is static, which performs the same operation on all input samples, hence it is not enough to deal with complex scenarios. We use the DynamicReLU function to encode the global up and down into a hyperfunction, and dynamically adjust the piecewise linear activation function instead of the Mish function accordingly, which solves the defect that the activation function cannot be dynamically adjusted, and realizes the detection of the headgear and seat of the Thangka image. The experimental results in Thangka image detection show that the evaluation protocol mAP of the algorithm reaches 37.9%, which is an increase of 3.8% compared to the YOLOv4 algorithm. The algorithm improves detection accuracy without increasing the depth and width of the network.