IMPROVEMENT OF YOLOV4 BASED ON BIDIRECTIONAL FEATURE PYRAMID

In order to further improve the target detection accuracy of YOLOv4 algorithm, an improved YOLOv4TB (YOLOV4+Transformer+BIFPN) algorithm is proposed. The proposed algorithm used Transformer to extract co-occurring object features on the same feature map for feature enhancement, and replaced the PAN model with BIFPN (bi-directional feature pyramid network) model. The problems of redundant calculation and the same contribution of different feature layers in YOLOv4 were solved. On this basis, the Leaky-ReLU activation function and the separable convolution technology were used to solve the problems of the decreased accuracy of target detection and the increase of the number of parameters and computation. Experimental results on Pascal VOC data sets show that compared with YOLOv4, YOLOv4TB algorithm has higher detection accuracy and operating efficiency, and the number of parameters and computation are reduced.