Research On Target Recognition In SAR Images

Synthetic Aperture Radar(SAR)can operate in harsh weather and weather conditions,with the continuous improvement of SAR resolution,the acquired images can be widely used in military and civilian fields,especially in military tactics in the normalization of the investigation and monitoring has important application value.Typical SAR automatic target recognition needs to go through three stages of target detection,target identification,target recognition.Compared to the target detection,target identification,SAR image target recognition there are still problems in the identification effectiveness and accuracy,therefore,SAR image target recognition research has become a difficult problem in the field of Information Engineering Research,which has important research significance.In this dissertation,the recognition accuracy of the existing target recognition method is not high under different configuration conditions(including pitch,occlusion,noise,etc),key techniques for Target Recognition in SAR Images: In-depth theoretical analysis and feasibility study are carried out on the method of target feature extraction and target classification,The specific research content is as follows:1.In order to solve the problem of image information loss when extracting the feature of SAR image based on traditional two-dimensional empirical mode decomposition,an improved SAR image target recognition method based on complex two-dimensional empirical modal decomposition feature extraction is proposed.The SAR image itself is a complex matrix with amplitude and phase information,the simple use of amplitude information in the traditional two-dimensional empirical modal decomposition method will cause a certain degree of lack of information.Complex two-dimensional empirical modal decomposition as an extension of traditional two-dimensional empirical modal decomposition in the complex domain,it can be effectively used for the processing and analysis of complex images.The original SAR image is decomposed by complex two-dimensional empirical mode decomposition to obtain a multilevel complex intrinsic mode function,reflecting the two-dimensional characteristics of the target time-frequency.Combined with sparse representation classification algorithm,intrinsic correlation constraints can be used to improve the characterization accuracy of complex intrinsic modulus functions at all levels.Based on the MSTAR data set,experiments are carried out under standard operating conditions and extended operating conditions,respectively,the experimental results show that this method can effectively improve the recognition accuracy.2.The characteristics obtained by a single random projection matrix designed based on compression perception theory for SAR images have limitations,the characteristics of the original SAR image cannot be reflected in many ways,an improved SAR image target recognition method based on joint multilevel two-dimensional compression perceptual projection feature extraction is proposed.Perception of random projection matrices by building multiple 2D compression which extract multi-level features of the original SAR image,multi-level two-dimensional compression-aware random projection matrix can not only effectively reduce the dimension of high-dimensional data processing,and the feature vector obtained has good discrimination,it is possible to reflect the different levels of characteristics of the original SARimage of the target,joint sparse representation is used to represent multiple features extracted,and the intrinsic relationship of different features is investigated,the target category is determined based on the sum of the reconstruction errors of all features.Based on the MSTAR data set,experiments were carried out under standard operating conditions and extended operating conditions,the experimental results show that the method has good recognition rate and robustness.3.In view of the redundancy problem of multi-feature vector Fusion when SAR image is extracted based on convolutional neural network,a deep feature fusion SAR image target recognition method based on multi-set typical correlation analysis is proposed.Through the convolutional operation using different convolutional cores in convolutional neural networks,the multi-layer depth characteristics can be excavated and learned,Feature vectors from different convolutional layers are fused into the final feature vector using multiple sets of typical correlation analysis,reduce redundancy while maintaining their relevance,sparse representation classification is used to classify the depth feature vectors of multiple sets of typical correlation analysis fusion,the experimental results show that the method maintains a high recognition rate and robustness under standard operating conditions and extended operating conditions.4.The SAR image target recognition method based on two-dimensional empirical modal decomposition and limit learning machine is proposed,two-dimensional empirical modal decomposition of SAR image feature extraction,resulting in a multi-layer two-dimensional internal mode function,these multi-layer two-dimensional internal model functions can effectively eliminate noise,it compensates for the relatively low adaptive performance of the limit learning machine.Extreme learning machine as a new type of fast learning algorithm,the training can be carried out without adjusting the parameters during the training process,when the number of neurons in the hidden layer is set,the optimal solution is automatically obtained,with high learning efficiency and strong generalization ability.Based on the MSTAR data set,the method has a high recognition rate and robustness under both standard operating conditions and extended operating conditions.