Research On Improving Target Detection Algorithm Based On YOLOv3

As one of the core research tasks of computer vision,object detection has been widely used in various fields,such as traffic detection,medical diagnosis,security and so on,and has been widely concerned by researchers.With the continuous development of artificial intelligence technology,target detection algorithm based on deep learning has become the mainstream algorithm of current target detection,among which,YOLOv3 algorithm is one of the most concerned algorithms among deep learning target detection algorithms,and has the advantages of high detection efficiency,strong generalization ability,and good robustness.However,with the increasing complexity of detection tasks,the detection accuracy of YOLOv3 algorithm can no longer meet the requirements of applications.In view of this,this paper studies the YOLOv3 algorithm and improves it in the main feature extraction network and candidate box screening stage to achieve rapid and accurate detection of complex target tasks.The main research contents of this paper are as follows:(1)Analyze the detection process of YOLOv3 algorithm.By introducing Res2 Net module into dark NET-53,the backbone network of YOLOv3 algorithm,and re-stratification and combination of feature information within a single residual block,a new residual structure is constructed,which makes the network make more full use of context information in feature extraction for extract finer-grained features.In the prediction box regression part,the method of reducing the confidence of the prediction box is used to replace the screening method in the original algorithm to reduce the probability of multiple target missed detection.At the same time,the penalty terms of distance,overlap rate and aspect ratio between two frames are added into the loss function of screening prediction frame to improve the positioning accuracy of prediction frame.Experiments on Pascal VOC data set show that the m AP value of the improved algorithm increases from 78.3% to 81.7%,which verifies the rationality of the improved scheme.(2)Two attention mechanism modules,SE module and CBAM module,are analyzed and integrated with the improved residual structure to construct a new attention residual module,which can improve the network’s attention to important information and improve the detection effect of the model.The existing data enhancement strategies are analyzed and the data set used in the experiment is processed to improve the generalization ability of the improved model.Pascal VOC data set was used for experimental analysis: after the introduction of attention mechanism and data enhancement,the detection effect of the algorithm was further improved,and m AP was increased from 81.7% to 83.6%,which verified the rationality of the improved scheme.