Research On SAR Image Classification Algorithm Based On Deep Feature Learning

Synthetic Aperture Radar(SAR)is an active microwave imaging system with all-day and all-weather imaging characteristics,which plays an important role in military and civilian fields.SAR imgae classification is the key to SAR image interpretation.However,the classification performance of traditional features extraction-based methods is limited due to the complex texture,intra-class diversity and inter-class similarity,and few-shot of SAR image.In recent years,with the development of deep learning,Convolutional Neural Network(CNN)has received extensive attention in the field of SAR image classification due to its powerful ability to extract deep features.Therefore,based on CNN,according to the characteristics of SAR images,this dissertation starts from the extraction of discriminative SAR image features by CNN and conducts research on SAR image classification methods based on deep feature learning.The main research work of this dissertation is as follows:(1)Aiming at the problem that the insufficient feature representation of complex texture area in SAR images,a SAR image classification method based on dense-connected covariance network is proposed.The method realizes the fusion of shallow low-level features and deep high-level features through dense-connected network structure,and enriches feature completeness.At the same time,the global covariance pooling is used to extract second-order statistical features,thereby enhancing the feature representation ability of complex texture area.The experimental results show that the method can effectively improve the classification accuracy of complex texture area in SAR images.(2)Aiming at the problem of insufficient feature separability of different objects in SAR images,a SAR image classification method combining attention mechanism and center loss function is proposed.The method uses the attention mechanism to dynamically adjust the weight information between the input features,strengthen the role of useful features on classification,and suppress the impact of invalid features on classification.On the basis of the traditional cross-entropy loss function,the center loss function is dedigned to improve the intra-class consistency,thereby enhancing the separability of features.The experimental results show that the method can effectively improve the classification accuracy of SAR images.(3)Aiming at the lack of effective features in the edge area of SAR images,a SAR image classification method combining superpixel segmentation and self-attention network is proposed.The method performs spatial weighting on the input image blocks on the basis of superpixel segmentation,so as to effectively suppress the interference information of edge area.On this basis,a multi-scale self-attention network is used to fuse the contextual information of the edge area to further enhance the image edge features.The experimental results verify the effectiveness of the method in SAR image edge localization.(4)Aiming at the problem of insufficient feature generalization ability under few-shot of SAR images,a SAR image classification method based on dual-scale siamese skip-connected network is proposed.The method utilizes the siamese network structure to improve the generalization ability of features by expanding the number of training samples and using a contrastive loss function.On this basis,a dual-scale siamese network is constructed to further optimal the classification performance of image homogeneous area and edge area.The experimental results verify the effectiveness of the method.