Research On SAR Image Target Recognition Based On Sparse Representation Tree

Automatic target recognition (ATR) plays a more and more important role in civil and military fields such as social security, environmental monitoring and homeland defense.Synthetic Aperture Radar (SAR), as an important part of modern target detection system, which has the advantage of breaking the light, weather, time limit and obtain high-resolution target images, becomes a reliable target identification data source.This thesis focus on the SAR image target recognition, especially the variant target recognition. A sparse-representation-tree target recognition method based on sparse rep-resentation and dictionary learning is proposed.Work in this thesis mainly carries on the following two parts:In the target recognition , the misclassification will arise when multiple variants are from the same target type in the target set to be identified. A Sparse Representation Tree(Tree-SR) framework for radar target recognition based on the sparse representation method is proposed to solve the problem that general methods likely obtain high target type recognition accuracy, but with low variant target recog-nition accuracy. The Tree-SR is divided into two levels: type identification and variant identification.The root node is built to identify target type, while child nodes of the tree are built to identify variant target series. Experimental results on the MSTAR dataset show that the Tree-SR method improves the recognition ability of the variant target by about 6% while the target recognition accuracy is competitive to the mainstream method.To solve the problem that Tree-SR method needs a specific type-variant partition in the target set,an improved Tree-SR is proposed. The improved Tree-SR(Improved-Tree-SR) uses a cluster method to automatically learn the type-variant partition so that the proposed method is more applicable. The root node is used to allocate input SAR images with similar sparse results to different children nodes, which have more specialized dictionaries and classifiers to identify target series. Experiments on the MSTAR target dataset show that the established tree structure is consistent with the sample distribution and improves the target recognition rate to about 84%. Comparing with the original sparse coding method,the tree-structure supervised classifier effectively improves the target series recognition accuracy while is still computationally efficient.