Research On First-order Target Detection Algorithm Based On Deep Learnin

As a research hotspot in the field of machine vision,object detection has been widely used in medical,military,security and other fields.Although the existing onestage object detection algorithms have excellent real-time performance,they have some shortcomings in fine-grained object detection and object boundary frame accuracy.It also has some shortcomings when applied in embedded devices with limited computing power.How to ensure the balance of detection accuracy and detection speed in the application of one-stage object detection algorithm becomes a problem that needs to be solved.This paper summarizes the research of object detection algorithm,algorithm framework and algorithm skills in recent years.Aiming at the problems in the algorithm and the actual application situation,the YOLO series algorithm in the onestage object detection algorithm is taken as the research object.The framework and loss function of the algorithm are improved to improve the accuracy and speed of target detection and reduce the complexity of the model.The main work of this paper is as follows:(1)Because existing one-stage object object algorithms can not achieve accurate detection of complex target tasks.Taking YOLOv3 algorithm as the optimization object,the activation function of the algorithm Leaky Re LU,border regression loss function Smooth L1,classification loss function cross entropy loss function and algorithm feature fusion network are optimized.Replace the Leaky Re LU activation function with the Mish and Swish activation functions,and the Smooth L1 border regression loss function with the GIOU and CIOU loss functions.The experimental results showed that m AP value of the algorithm’s object detection accuracy improved the most when Mish activation function and CIOU loss function were used,from77.44% to 78.67%.Using the idea of Focalloss,α and γ parameters were used to control the weight values of positive and negative samples and difficult samples in the cross entropy classification loss,respectively.The results showed that when α=0.75 and γ=2,the m AP value of the algorithm improved the most,from 78.67% to 80.00%.Spatial Pyramid Pooling(SPP)and Convolution Block Attention Module are added to feature fusion networks.CBAM increases the m AP value of YOLOv3 algorithm from80.00% to 82.25%.Compared with the YOLOv3 algorithm before optimization,the optimization method in this paper improves the m AP value of the target detection accuracy of the YOLOv3 algorithm from 77.44% to 82.25%,with a total increase of4.81%,showing obvious optimization effect.(2)Aiming at the problem of slow object detection under embedded devices,this paper designs a lightweight object detection model based on YOLOv4 algorithm model.The feature extraction network CSPDark Net-53 and the ordinary convolution and SPP modules in the feature fusion network of YOLOv4 algorithm are optimized.Using Mobile Netv3 and Ghost Net network replace the CSPDark Net53 network of YOLOv4,and the common convolution in feature fusion network is replaced by deep separable convolution.The appeal method is to reduce the computational amount of YOLOv4 algorithm.The ASPP module combined with deep separable convolution was used to replace the SPP module in the feature fusion network of the YOLOv4 algorithm,and the Efficient Channel Attention(ECA)module was added into the three output channels of the feature extraction network.The results showed that the computational load of the optimized YOLOv4 algorithm decreased from 64,363,101 to 17,013,068,which was reduced to 26.4% of the original YOLOv4 algorithm.The object detection accuracy m AP decreased slightly,from 71.56% to 65.95%.Through the above improved method,the computational load of YOLOv4 algorithm is greatly reduced,and the accuracy of target detection is guaranteed.