Weight Optimization for Multi-Task Sparse Representation in Sar Image Target Recognition

Gabor wavelets filters with different orientations and scales were applied on SAR image as a feature extraction technique. However, due to the different characteristics of the constructed Gabor filters, different Gabor features could have different impact on material representation, influencing the recognition rate eventually. In this paper, a novel Gabor weight optimization based multi-task sparse representation is proposed for synthetic aperture radar (SAR) image target recognition. First, each Gabor feature is sparsely represented over the corresponding set of Gabor features of all training samples under multi-task sparse representation framework. Then, the weights of multi-task representation are optimized by a least-squares optimization with l2-norm regularization according to the loss function defined by the classification results of the classifiers. The final classification results are acquired by a weighted fusion strategy. Experiment results prove the effectiveness of the multi-task sparse representation method based on weight optimization.