Deep Learning For SAR Image Target Recognition And Terrain Classification

Originated in artificial neural networks, imitating human brain computing model, deep learning is widely used in image field, especially target recognition and classification. With the advances in remote sensing, SAR image occupies an important position in military and civilian fields with its large amount of information, all-weather all-time characteristics. For both tasks of SAR image, feature determines the upper limit. Deep learning can automatically learn more abstract features from the raw input. Among deep models, deep belief network, DBN) is a generative model containing data characteristics. DBN is composed of stacked RBM for unsupervised learning and supervised back-propagation for fine-tuning, thus learning more suitable features for recognition and classification. In this paper, DBN is used for SAR images. Details are as follows:First, deep learning for SAR target recognition. Due to the limitation of MSTAR data, the convergence of deep learning is not easy. We proposed the combination of data fusion and deep model strategy, extracting contourlet and curvelet features combined with raw data as input of DBN, along with gaussian RBM for target recognition. We get better accuracy. Unlike RBM, convolution RBM(CRBM) takes 2-D structure and spatial information of SAR image into account. And each kind of weight corresponds to a feature operator, thus extracting shared feature of different natures. We proved CRBM can get better accuracy.Second, deep learning for POLSAR terrain Classification. Traditional RBM is more suitable for binary data and can expand to data of exponential family. Therefore, to meet the real POLSAR data, we applied gaussian RBM conducting DBN; to meet POLSAR complex data,we proposed wishart RBM conducting DBN. Specific steps are as follows: we use the covariance matrix of POLSAR data as input, the stacked wishart RBM for pre-training, coupled with the back-propagation to fine-tune the model,and perform the final classification using softmax classifier. Classification accuracy has been improved.