SAR Image Target Recognition Based On Sparse Representation And Deep Learning

The Synthetic Aperture Radar (SAR) is a form of active coherent imaging radar, which is applied in many areas, such as environmental monitoring, resource exploration, national military and so on. The detection and recognition of a specific target using SAR is the most important applications in the military field. How to achieve automatic target recognition (ATR) with SAR images, has important theoretical and practical significance. Thus SAR image target recognition becomes a research hotspot worldwide. To overcome the problems caused by the speckle noise and non-robust low-level features, we propose two SAR image target recognition algorithms based on multi-scale sparse representation and deep stacked denoising autoencoder network, which take advantage of the good feature learning of sparse representation and high-level feature extracting of deep learning. The main work as follows:1. Multi-scale sparse representation based SAR image target recognition approach. Traditional sparse approach for SAR target recognition usually ignores scale information. Therefore, we propose a multi-scale sparse representation algorithm based on dense SIFT features. It firstly extracts the multi-scale dense SIFT features, then uses them to train a multi-scale dictionary by dictionary learning algorithm. After obtaining the trained dictionary, multi-scale sparse representation of the SIFT features is solved by sparse coding algorithm. Finally, the algorithm trains a linear SVM classifier to finish recognition with the multi-scale sparse features.2. Deep denoising autoencoder network based SAR image target recognition approach. Considering the speckle noise in SAR images and the robustness of denoising autoencoder, we propose a deep stacked denoising autoencoder network based SAR image target recognition approach. First of all, we still extract a lot of dense SIFT features as inputs to train the network which is hardly overfitting and more robust to noise. With the trained deep network, it can generate effective high-level feature representation to train a classifier.In this thesis, according to the basic theory of sparse representation and deep learning, we use multi-scale sparse representation and deep denoising autoencoder network to build the framework of feature learning, then we propose two novel SAR image target recognition algorithms. The MSTAR database and ship database collected from TerraSAR-X images are used in classification setup. Results show that both algorithms can learn more effective and robust feature representations and improve the performance of SAR image target recognition.