SAR Images Classification Method Based On Deep Sparse Network

The Synthetic Aperture Radar(SAR) image contains abundant information. It is not affected by bad weather and night. More importantly, it can work all-day and all-weather to achieve the purpose of monitoring land and oceans. The classification of SAR images is widely used in both civil and military. However, the SAR image has a large number of speckle noise, and the traditional method is easily affected by speckle noise. At the same time, the target of SAR is complex. For different scenes, the feature extraction methods of different resolution SAR images are time-consuming and laborious, and their generalization ability are poor.Recently, deep learning has become a hot research topic at home and abroad, and it has many successful applications in image processing, signal processing, computer vision and patter recognition. In this paper, the sparse features of SAR images are excavated, and the objects of the SAR images are classified by using deep sparse network models which overcome the effects of speckle noise, have strong generalization ability and are simple designed. Specific research contents are as follows:(1) A SAR image classification method based on hierarchical sparse auto-encoder convolutional neural network is proposed. Firstly, the SAR image blocks are used to train the first layer sparse auto-encoder model. Secondly, by the convolution operation, the first layer sparse auto-encoder model is used for the whole SAR image. After the extraction of the sparse features of the first layer, we train the second layer of the sparse coding model, and with the help of convolution operation, we extract the sparse feature of the second layer, then the sparse feature of the second layer is feed to Softmax classifier. The Softmax classifier is trained with labeled samples. Finally, the labeled training samples are used for weights fine-tuning of the neural network. And we get the hierarchical sparse auto-encoder convolution neural network. The performance of the proposed algorithm is verified in the SAR image database which is composed of various ground objects. The experimental results show that the proposed algorithm can effectively improve the classification accuracy of SAR images. Even in small sample learning situations, it can still get higher classification results.(2) A SAR image classification method based on hierarchical sparse filtering convolutional neural network is proposed. The feature extraction of the algorithm is an unsupervised process. The sparse filter is trained by an unsupervised method, and is used to extract the feature of a large SAR image. The classification of the algorithm is part of a supervised process. A combination of sparse features extracted from the first layer and the second layer sparse filtering network is used for the training of linear kernel SVM combined with the labeled information of the samples. The performance of the proposed algorithm is verified in the SAR image database. The results show that: compared with the traditional convolution neural network, the feature extraction algorithm of the algorithm is not affected by the number of training samples, its training efficiency is high. Even in the case of small sample learning, the classification result of the proposed algorithm is still better than the traditional methods.(3) A SAR image classification method based on hierarchical 3D sparse filtering convolutional neural network is proposed. The algorithm is combined with the idea of convolutional neural network. The feature extraction of 3D sparse filter is applied to the feature extraction of large SAR images. The training process of the sparse network of each layer is carried out in an unsupervised manner, and then be fine-tuned layer by layer. Finally, the parameters of the feature extraction network are determined by the way of global fine-tuning. The classifier is Softmax. The performance of the algorithm is verified in the SAR image database consisting of a multi class SAR objects.(4) A SAR image classification method based on hierarchical 3D sparse filtering NIN convolutional neural network is proposed. Apart from the 3D sparse filtering, the NIN network structure is added to the local threshold, and the nonlinear transformation is introduced to enhance the discrimination of the model. At the same time, the global averaging pooling of NIN network can enhance the connection between the feature map and the category, and it is not necessary to learn the parameters. The performance of the proposed algorithm is verified by the SAR image database which is composed of a variety of ground objects. The experimental results show that the proposed algorithm can effectively improve the classification accuracy of SAR images.