Detection Of Small And Medium Targets Based On Improved SSD Model

With the rise and rapid development of artificial intelligence in recent years,the research methods of deep learning are gradually deeply integrated with some traditional research directions in the computer field,and have yielded a great number of positive results.Computer vision is the first field where deep learning has made breakthroughs.With the development of computer hardware,the improvement of computer hardware capability and the reduction of cost also make the computing power required by deep learning accessible.However,in the field of computer vision,there are constantly various applications of convolutional neural network on target recognition.From the earliest Alex Net model to the current end-to-end SSD network model,the research on target recognition has made rapid development.The target detection technology based on deep learning has gradually replaced the traditional target detection technology due to its reliable advantages,namely higher accuracy,stronger robustness and real-time performance.In the face of new technological changes,however,there are new challenges,with the emergence of complex scenes,because most of deep learning training data sets are based on an existing data set for training and prediction,so often in according to the characteristics of the small and medium target of information to predict when they omit many details characteristics,resulting in relevant network model testing capacity constrained.This becomes a key problem when upgrading and optimizing network models.Therefore,this,this thesis analyses the issue:(1)After studying and analyzing the classic target detection network,this thesis selects the most popular multi-scale convolution feature detection network model,namely SSD network model,for optimization.After analyzing the convolutional layers of the SSD network model and the feature diagrams of each layer,in view of the insufficient detection ability of detecting small targets,the idea of connecting context information was adopted for the shallow network,that is,the network structure of Conv4 was improved by using the method based on the residual network idea.In addition,aiming at the problem of insufficient semantic information richness of the shallow convolutional layer,feature fusion was conducted for the feature graphs of Conv4＿3 and Conv5＿3,a new SSD network model was designed,and the feature graphs output from the improved shallow convolutional layer were used for the target detection of theconvolutional detection layer.The experimental results show that the detection effect of the modifield model SSD＿Fusion V1 is significantly improved in the detection of small and medium targets.(2)On the basis of the first part of the research,aiming at the problem of insufficient performance of Conv7 in the additional convolutional layer of SSD network model with respect to the detailed information,feature fusion of Conv3＿3 feature map was conducted to enhance the performance of the feature map of the layer with respect to the small and medium target details;Target feature region mapping magnification method is adopted for the target recommendation region output by the SSD network model after two improvements,so that the target feature information output by the improved model can be mapped to a new target region information feature map with more details.The experimental results show that the final model SSD＿Fusion V2,after two improvements and using the regional mapping amplification mechanism,the detection effect has significantly improved compared with other classical algorithms on the detection effect of specific categories in this thesis and on the detection of small and medium target objects and on the public number set compared with the original model SSD.Starting with the SSD based network Conv4 convolution unit network structure,Conv4 Conv4＿3characteristics of shallow layer convolution output figure,Conv7 convolution layer output characteristic figure,there are three methods for convolution layer to improve low performance of detecting target ability.The methods respectively are feature fusion method 、reference to residual network structure、method of using the target area feature map zoom.