Synthetic Aperture Radar(SAR)has been widely used in military and civil fields because of the ability of all-day and all-weather imaging and that of continuous observation.With the development of SAR sensor and imaging technology,the resolution of SAR images is getting higher and higher.Furthermore,with the overwhelming amount of SAR images available,there is a huge demand for SAR automatic target recognition(ATR).The most popular SAR ATR system consists of three stages: target detection,target discrimination and target recognition.Among them,target detection aims to find targets from large scene SAR images and then extract target chips,and target recognition aims to classify the types of target chips.These two stages are widely concerned and studied by scholars at home and abroad.Whether it is target detection or target recognition,effective target feature representation is an extremely important factor.However,traditional SAR target detection and recognition methods usually only use a single feature to describe the target characteristics,which is difficult to achieve complete and sufficient target representation,and then,limits the performance of target detection and recognition to a certain extent.Combined with the idea of feature fusion,that is,mining and fusing complementary features to fully characterize and describe the target information,the dissertation mainly studies the related theoretical and technical problems of SAR target detection and target recognition.The main contents of the dissertation can be summarized as follows:1.Aiming at that avoiding the clutter statistical modeling in the traditional Constant False Alarm Rate(CFAR)target detection method and taking advantage of clutter information,we carry out the research that introduces the saliency detection method based on sparse reconstruction into SAR target detection task.With feature fusion,a SAR target detection method based on dual domain sparse reconstruction saliency fusion is proposed.Although the traditional saliency detection method based on sparse reconstruction can make full use of clutter information and avoid clutter statistical modeling,it only uses a single amplitude feature,which is easily affected by the speckle noise and has weak representation ability for SAR target.Therefore,the proposed method extracts pixel-level amplitude features in the image domain and superpixel-level structural features in the structure domain,which describe the target from different angles and have a certain degree of complementarity.The two different features are fused to achieve a more complete and sufficient characterization of the target characteristics and a more robust and accuracy saliency map which can highlight targets and suppress clutter,so as to improve the performance of SAR target detection in complex scenes.Experimental results based on measured mini SAR data show that the proposed method has better target detection performance than traditional SAR target detection methods and other saliency detection methods.2.Aiming at that avoiding the clutter statistical modeling in the traditional CFAR target detection method and taking advantage of the representation ability in convolution neural network for improving the target detection performance in complex scenes,we carry out the research that introduces the Single Shot multibox Detector(SSD)based on convolutional neural network into SAR target detection task.With feature fusion,a saliency-guided SSD target detection network for SAR target detection is proposed.Although SSD target detection network has strong representation ability drawn support from convolution neural network and can avoid clutter statistical modeling,it adopts indifference processing in feature extraction on scene data,and only uses the current scale features for target detection on a single scale,which limits the target representation ability.Therefore,on the one hand,the proposed method introduces saliency feature into SSD target detection network by the way of fusing saliency feature with network feature,which the ability of enhancing target regions and suppressing clutter in saliency feature is used to enhance the distinction between target and clutter.On the other hand,dense connection structure is used on SSD target detection network to fuse feature information of adjacent scales from bottom to top,which mines the correlated complementary advantages between adjacent scales and introduces context information to enhance the representation ability of target,so as to improve the performance of SAR target detection in complex scenes.Experimental results based on measured mini SAR data show that the proposed method has better target detection performance than the traditional SAR target detection method and other target detection methods based on deep learning.3.Although the bottom-up feature fusion based on dense connection is beneficial to the performance improvement of SSD target detection network,its performance improvement is still limited by the one-way information propagation during feature fusion.Therefore,modifying the fusion method of different scale features,a cross scale feature fusion SAR target detection network with rectangle-invariant rotatable convolution is proposed.The proposed method introduces a cross scale feature fusion module in SSD target detection network.The cross scale feature fusion module performs bottom-up and top-down two information flow to propagate the information in different scale features,achieving the fusion of related complementary information between cross scale features,which further enhances the representation ability of target,and then improves the performance of target detection.Meanwhile,considering the inherent mechanism of fixed sampling points of conventional convolution layer in SSD limits the adaptability of target geometric transformation,a rectangle-invariant rotatable convolution is proposed with the prior shape information that the vehicle targets in SAR images remains the rectangular shape on the basis of conventional convolution,making the sampling points can be adjusted adaptively according to the target.Rectangle invariant rotatable convolution is introduced into SSD target detection network to enhance the ability of representation of SAR target and modeling of geometric transformation.Experimental results based on measured mini SAR data show that the proposed method has better performance than the traditional SAR target detection method and other target detection methods based on deep learning.4.Aiming at that most traditional SAR target recognition methods only use single shallow feature to represent the target,we carry out a SAR target recognition algorithm with feature fusion and deep feature mining,and propose a multi-level feature fusion SAR target recognition method based on deep forest.The proposed method extracts the amplitude features and structural features that describe the target from different angles and have a certain degree of complementarity.Based on the amplitude features and structural features,the multi-level amplitude features and multi-level structural features are further extracted,which can describe the target from local to global,making a more complete and sufficient characterization of the target characteristics.Based on the deep forest model,multi-level amplitude features and multi-level structure features are fused through stacked multiple forests,and the deep information of features can be mined through layer by layer nonlinear mapping,which enhance the ability of target representation and improve the performance of SAR target recognition.Experimental results based on MSTAR data show that the proposed method has higher recognition accuracy than traditional SAR target recognition methods and some SAR target recognition methods based on deep model. |