An Improved Target Detection General Framework Based on Yolov4

Abstract

The speed and precision of the target detection algorithm have received wide attention in the application. This paper proposes an improved network framework for the yolov4 algorithm, which reducing the parameters of the model without lowering too much accuracy. Firstly, the up-sampling and down-sampling links of the PANet are strengthened with an increased CBAM attention mechanism, which improve the algorithm ability to deal with the object occlusion. The depth separable convolution is introduced to reduce the amount of model parameters and improves the algorithm speed. Secondly, the Se-Net attention mechanism is added to the residual module of CSPDarknet53 to pay more attention to the channel. Thirdly, Soft-NMS is used to optimize the screening of the detection frame during the detection process. Experiments show that the comparison of VOC2007 dataset is 84.26%, and in the VOC2007+VOC2012 dataset is 91.2%. The detection speed of FPS has increased by 2.21.