Research On Small Object Detection Algorithm Based On Deep Learning

Small object detection refers to the positioning and recognition of low pixel or low proportion of the image object.It has important application value in unmanned driving technology,intelligent traffic monitoring,computer-aided diagnosis technology and so on.Because the small object occupies less pixels in the image,the information reflected in the image is less and the background interference is greater,using the general object detection method to detect the small object is easy to cause the small object to miss and misdetection.The research purpose of this thesis is to improve the detection accuracy of small objects on the basis of the existing object detection algorithm,and reduce the detection rate of small objects by means of structural adjustment,optimization and increasing the semantic information of small objects.Combining with the distribution of the object scale in the data set,this thesis improves the existing object detection algorithm from three aspects: the processing process of the data set,the structure of the algorithm network and the selection of the prediction box.In the network structure,the dense convolutional network is used to replace the VGG network as the feature extraction network,and the depth of the dense convolutional block is set according to the principle of effective sensing field and size alignment of anchor frame.At the same time,the prediction network uses cascade to fuse the deep and shallow features of the image,so as to form a feature fusion network structure.For the implementation strategy of the algorithm,on the one hand,the method of repeated sampling of small object images and augmentation of replication of small object images are adopted to conduct data augmentation of small object in the data set.On the other hand,in the process of Non-Maximum Suppression,the measures of reward and punishment are added respectively to solve the problem that the score of prediction box and the influence of center distance of anchor box are not fully considered in the existing methods.In this thesis,VOC and self-constructed wheat spider data sets were used for comparison.The results show that the detection accuracy of small objects in two data sets can be improved by 3.63% and 2.24% respectively without reducing the detection accuracy of common objects by using the feature fusion network structure with increased context.The improved NMS method was used to improve the detection accuracy of small objects in VOC data set and wheat spider data set by 2.84% and 2.13%,respectively.In addition,when the number of samples of small targets in the data set is insufficient,the method of data augmentation can significantly improve the detection accuracy of small objects.Experiments and analysis show that the proposed improvement strategy can well solve the problem of insufficient number of small objects,lack of semantic information and easy to be suppressed in the NMS process,so as to greatly improve the detection accuracy of small objects.It can meet the requirements of conventional object detection applications.