Research On SSD-based Multi-scale Detection Algorithm

In recent years,with the rapid improvement of high-intensity and fast-paced full information operation level,the detection of vehicles and other military targets depends on the traditional manual detection method,which puts forward higher requirements and challenges to human and material resources,and it is difficult to guarantee the real-time and accuracy of the detection results of the input image.With the continuous improvement of military automation,the automatic detection of multi-scale vehicle targets and other important military targets by using the target detection technology based on deep learning becomes an inevitable trend in the future development of related military target detection.Therefore,this paper explores the multi-scale target detection algorithm of vehicles based on SSD(Single Shot Multi Box Detector),the main work is as follows:First of all,the convolution neural network is studied,including convolution,pooling,activation function,it focuses on the basic principle of SSD based target detection algorithm,The data set of vehicle is made independently,and the multi-scale target detection of vehicle based on VGG basic network is realized through the built Pytorch deep learning platform,and compares the influence of VGG and Res Net on algorithm accuracy and speed,according to the characteristics of deep network model degradation in VGG basic network,and finally selects Res Net18 as the basic network of SSD algorithm in this paper.Through analyzing the main difficulties of multi-scale vehicle target detection,the multi-scale SSD target detection algorithm defects in target detection of vehicles is found out.Then,aiming at the defect of SSD algorithm in small target of multi-scale vehicle target,the algorithm based on SSD is improved.Based on the K-means clustering results of default boxes scale,the default boxes length width ratio of vehicle target is designed,which is more suitable for the default boxes size of vehicle data set.It solves the problem of unreasonable design of initial suggestion frame for small target in SSD network,and improves the detection performance of small target of vehicle.In view of the fact that the low-level feature map contains small target information but lacks corresponding semantic information,the high-level feature map contains richer semantic information but lacks small target information,this paper proposes to improve the network model based on the fusion of the features of each layer of FPN to complete the feature fusion between low and high layers,so as to improve the detection ability of small targets of vehicles.Finally,the fusion conv4＿3 feature map is used as the bottom feature map of the feature pyramid.Then,C4＿3 is integrated into fc7,C6＿2,C7＿2,C8＿2 and C9＿2 feature maps to improve the detection ability of small targets,and is verified by experiments on the self-made vehicle data set and voc207 data set to achieve the improvement of network model performance.Finally,the deep learning experimental platform is built for Pytorch.On the basis of the platform,the visual operation interface based on multi-scale vehicle target detection is completed.