Research On Target Detection And Recognition In SAR Remote Sensing Image Based On Deep Convolutional Neural Network

SAR(Synthetic Aperture Radar)is a useful tool for detecting and recognizing targets.SAR’s unique imaging method makes it widely used in remote sensing target detection and recognition fields and plays an extremely important role.Nowadays,with the continuous development of science and technology,the performance of SAR continues to improve,which makes more and more high-quality SAR remote sensing images available for practical applications.However,target detection and recognition based on SAR remote sensing images are more difficult due to the inherent speckle noise,ambiguity,and instability of target images.Therefore,the research on target detection and recognition in SAR remote sensing images has important theoretical research significance.In order to promote the application of SAR remote sensing images in target detection and recognition,this study firstly analysed the characteristics of SAR remote sensing images and made use of deep learning to carry out research on speckle noise suppression,target detection,target recognition and typical target monitoring methods for SAR remote sensing images.The following are the main points of this thesis:(1)In order to solve the problem that the local edge information of traditional SAR image denoising model is easy to be mistaken for noise and discarded,a speckle noise suppression method for SAR remote sensing images based on residual optimization network was proposed.This method used the residual optimization strategy to build a speckle noise residual image learning model and obtained the optimal identity mapping of the speckle noise,thereby realizing the effective suppression of the speckle noise.The experimental results on simulated dataset,real dataset and target detection method showed that the model can not only effectively suppress speckle noise but also effectively protect the details of the local edges and textures of SAR image.The details of preservation and noise suppression were significantly improved.It also lays a good foundation for subsequent target detection,target recognition and typical target monitoring.(2)To solve the problems of low model detection accuracy caused by the large feature difference of multi-scale SAR targets in complex scenes,a multi-scale SAR target detection model in complex scenes was proposed.This method first adopted the Receptive Field Block(RFB)model to simulate human vision and strengthen the multi-scale feature extraction ability.Then,combined with the feature fusion method,the context information is effectively used to realize the asymptotic scale information enhancement of the multi-scale target features,which improves the representation ability of the model to the multi-scale SAR target features,and then solves the problem of poor detection accuracy of multi-scale SAR targets in complex scenes.The effectiveness of the proposed method was verified by the experimental results on two target detection datasets of Sar-Ship-dataset and SSDD.It also showed that this method can not only improve the feature representation ability of the model,but also effectively improve the multi-scale SAR target detection accuracy in complex scenes.(3)To solve the problem of the sample diversity of small sample SAR target recognition is not enough to reflect the real data distribution and the shallow model is prone to overfitting or underfitting,a small sample SAR target recognition model based on lightweight network is proposed.This method first adopts the attention mechanism to effectively increase the feature expression ability of the model without significantly increasing the parameters.Then,the attention feature selection method is used to effectively compress unimportant feature information and reduce the feature dimension,so as to realize the joint optimization of the detection accuracy and speed of small sample SAR target recognition.The effectiveness of the proposed method was verified by the experimental results on two small sample target recognition datasets of MSTAR and FUSARShip.It also showed that the proposed model can not only identify targets quickly and accurately,but also can be used for different types of SAR target recognition,no matter land targets or sea targets.(4)In order to solve the problem that the sea ice target monitoring algorithm is unstable in tracking the rotating sea ice and the tracking rate is sparse,a sea ice targets monitoring method in SAR remote sensing images based on feature tracking and pattern recognition is proposed.This method takes full advantage of the fast detection speed of feature tracking and the accurate detection of pattern recognition,and constructs a fusion matching method of feature tracking and pattern recognition.An efficient sea ice target monitoring method is realized,which effectively improves the robustness of sea ice target monitoring algorithm.Through a series of experimental tests and accuracy verification of buoy data,it is proved that the proposed method has a good capability of sea ice target monitoring,and can be effectively used in the quantitative analysis of sea ice target monitoring.In summary,based on the in-depth study of SAR image characteristics,this thesis makes full use of remote sensing image processing and deep learning technology to effectively suppress the speckle noise of SAR image,improve the performance of SAR image target detection and recognition,and successfully realize sea ice target monitoring based on SAR image.It can provide methodology and technical route support for SAR remote sensing intelligent interpretation.