Research On Target Detection Model Optimization Based On YOLOv3

In recent years,the technology of artificial intelligence represented by convolutional neural network has promoted the development of computer vision and its practical application.The field of target detection has also made great progress due to the successful application of CNN model.Various advanced target detectors emerge in an endless stream,and the application of computer vision based on target detection technology is also developing rapidly.The phase one target detection algorithm YOLOv3 combines the essence of YOLOv1 and YOLOv2,and improves the prediction accuracy under the premise of maintaining the speed advantage,especially enhancing the recognition ability of small objects.The main improvements are: adjusting the network structure;using multi-scale features for object detection;object classification using logistic instead of softmax.However,the existing target detector based on yolov3 also faces many problems and challenges.Firstly,the existing target detector based on yolov3 has some shortcomings in feature fusion.Secondly,the balance between the object parameters and the prior frame parameters of the target detector based on yolov3 is also deficient.In view of the above two problems,this paper studies as follows:First,this paper analyzes the shortcomings of the existing target detector based on yolv3 in feature fusion,and proposes a target detection network based on feature pyramid structure yolv3 +.On the basis of darknet-53,the next sampling(red box)is carried out additionally.At the same time,in the Yolo layer,using the FPN like network structure,the features are extracted from the higher-level feature map(7 × 7),and the sampling operation is carried out.Based on this improvement,the Yo layer is expanded from the original three layers to four layers,and the output is also changed into four layers.A large number of experiments show that compared with the previous target detector based on yolv3 +,the target detector based on yolv3 + can effectively improve the performance of target detection.Secondly,this paper analyzes the problem of the lack of balance between the object parameters and the prior frame parameters of the existing object detector based on lyov3,and puts forward two optimization ideas: first,the size of the prior frame has a targeted change to the data set,when the prior frame is redistributed,the larger prior frame is allocated to the higher-order feature map;second,the object category and image resolution are two aspects To eliminate the influence of object parameters on the detection results as much as possible.The experimental results show that the optimization based on the combination of object parameters and prior frame parameters can effectively improve the detection accuracy while ensuring the speed.