Research On Image Small Target Detection Method Based On Deep Learning

In recent years,the satellite remote sensing technology based on computer vision technology has developed rapidly.In the civil field,high-precision target detection is helpful to assist traffic management and urban planning.In the military field,high-precision target detection helps to precisely target hostile intrusions and hazards and maintain national security.However,high resolution and small target size are the key and difficult points in the target detection of high-resolution remote sensing images.The application of deep convolutional neural network has solved the problem of image feature extraction in traditional target detection and greatly improves the target detection accuracy.However,both the one-stage target detection algorithm represented by YOLO(You Only Look Once)and the two-stage target detection algorithm represented by R-CNN(Region-based Convolutional Neural Networks)need to clip or interpolate high-resolution remote sensing images to fit the size range.Clipping results in the loss of target context information at the edge of clipping after feature extraction,and overclipping consumes more detection time.However,when the image is interpolated and reduced to a fixed size,the small and medium targets in the original image are smaller through the interpolation size,and the feature map obtained through the deep feature extraction network disappears,which causes great difficulties in the detection of small targets.In view of the above problems,this paper proposes a high resolution small target detection algorithm suitable for arbitrary size by combining with deep convolutional neural network.The main work of this paper is as follows:1.In this paper,the basic concepts of convolutional neural network and the related technical principles of target detection are deeply studied,and several classical target detection algorithms based on convolutional neural network are analyzed in recent years.Finally,YOLOv3 network is used as the basic network architecture of this paper.2.By annotating and cleaning the collected satellite remote sensing images and combining with the selected data samples from the public data set,the database required by this paper was established and the preset Anchors of the remote sensing automobile data set in this paper were calculated by the optimized K-means++algorithm.3.Because YOLOv3 network because of the limitation of the downsampling can only input network on fixed input image size.This paper puts forward a kind of applied to any size of image small target detection network,the network on the basis of YOLOv3 network,using characteristic diagram in figure splicing structure patch,make to the original image input to the network,test phase characteristic of the figure is a good way to network with the dimension of feature extraction and feature map feature fusion,made up for the inadequacy of YOLOv3,preprocessing and postprocessing operation,reduced network can adapt to any dimension of the input images,at the same time,The detection network on the larger feature map is added to further improve the network's detection capability for small targets and ultra-small targets.The detection time of large-size image is shortened by avoiding over-clipping while ensuring the detection accuracy.4.Since the target background of remote sensing automobile data set is complex,and there is a dense scene of automobile targets,it is easy to appear the imbalance of positive and negative samples in the training process.In this paper,focal loss is used to replace the original cross entropy loss and reduce the weight of simple negative samples in the training.At the same time,considering that there is only one detection category,a weighted factor is added to balance the category loss and border regression loss in the loss function to make the network training process more concentrated optimization of border loss,and improve the detection performance of the model against single category target.In this paper,deep convolution neural network to detect small targets for arbitrary size image,is a kind of end-to-end target detection algorithm.The precision loss caused by the network's preprocessing of the original image is reduced.The loss function is optimized for the remote sensing vehicle data sets in this paper,mAP was improved by 14%on the basis of YOLOv3 network,and the detection time of images was shortened.