The Research On Target Detection Of SAR Image Based On Semi-Supervised Convolutional Neural Network

Synthetic Aperture Radar(SAR)is an active microwave remote sensor,which is of great significance in the fields of national economy and military.With the rapid development of SAR technology,it is of significant research value to quickly and accurately detect the targets of interest from SAR images.In recent years,with the development of convolutional neural network(CNN),the feature extraction capability of CNN is getting better and better.The detection results of SAR target detection methods based on CNN can achieve better detection results than traditional methods.However,there are some problems of those methods:1)CNN based detection methods require a large number of slice-level labeled training samples,but it takes a lot of manpower and material resources to label the SAR images at slice-level,which is very difficult.2)There are a lot of clutters in SAR images of complex scenes,which will adversely affect the detection results.This thesis focuses on the studies of SAR target detection for these two problems.The main works of this thesis are as follows: 1.Aiming at that the current CNN based detection methods require a large number of slice-level labeled training samples,in chapter 3,a SAR target detection method based on self-training is proposed.The image-level label simply marks whether the image contains a target,which is easier to label than the slice-level label.Therefore,using a small number of slice-level labeled samples and a large number of image-level labeled samples,the network can be well trained through semi-supervised learning algorithm.The proposed method trains the network through semi-supervised self-training method.Firstly,the target detection network is trained using the slice-level labeled samples.After training convergence,the initial candidate region set is generated through the output of the region proposal network.Then,the image-level labeled clutter samples are input into the network.The negative slices are selected from the output of the network and added to the candidate region set.Next,the image-level labeled target samples are input into the network as well,and then the selected slices of the output are added to the candidate region set.Finally,the detection network is trained using the updated candidate region set.The processes of updating candidate region set and training detection network alternate until convergence.The experimental results based on the measured SAR dataset demonstrate that,the proposed method can achieve better performance when using a small number of slice-level labeled training samples and a large number of image-level labeled training samples,and the performance of the proposed method is similar to the fully supervised training method using a much larger set of slice-level samples.2.Aiming at that the detection could be adversely affected by the natural and artificial clutters in the SAR images of complex scenes,and the CNN based detection methods rely on a large number of slice-level labeled training samples,in Chapter 4,a SAR target detection method based on semi-supervised and attention mechanism is proposed.The proposed method includes two network branches,including detecting network branch and scene recognition network branch.During the training process,slice-level labeled SAR images are input to detect network branch,and image-level labeled SAR images are input to scene recognition network branch.The loss function of the network consists of the two branch's loss functions.The backbone network of the two network branches are shared.In addition,the proposed method introduces the attention mechanisms to the shared part of the two network branches,so that target area can be enhanced and the clutter area can be suppressed automatically through the learned attention map.The network can focus on the target area.Therefore,the adverse effect of clutter on detection performance can be reduced,and the number of false detections and missed targets can be reduced.The experimental results based on measured SAR dataset indicate that,the proposed method can achieve better performance than the existing target detection methods with a small number of slice-level labeled training samples and a large number of image-level labeled training samples.