Target detection algorithm for improving feature fusion and global perception based on YOLOv4
The YOLOv4 algorithm has a good balance in detection speed and accuracy,but there are still drawbacks of inaccurate positioning frame and low detection rate,especially for small detection targets and great changes in scale.Dealing with these problems,a new YOLOv4-based target detection algorithm is developed.The algorithm utilizes an enhanced feature fusion module—PANet combined with the bidirectional feature pyramid network instead of PANet to increase cross-scale connections,introduce weights at the output to improve the expressiveness of important features and solve accuracy degradation as a result of multiscale changes.Afterward,a new global association network is adopted to improve the output of the Sigmoid function while reducing the average pooling and computation,strengthen the model's learning of the contextual relationship of the target,and reduce noise interference and global information loss.The RS-OD and NWPU VHR-10 datasets are employed here,with average detection accuracies being enhanced by about 5%and 3%,respectively;the generalization experiment uses the VOC2007 + 2012 public dataset,with the average detec-tion accuracy being enhanced by about 0.6%.The experimental results reveal that the improved algorithm can effect-ively enhance the detection ability of the model.
YOLOv4target detectionfeature fusioncross-scalemultiscale variationglobal attentionaverage pool-ingcontextual information
国家科技期刊平台
NSTL
万方数据
