Hyperspectral target detection via locality-constrained group sparse representation

Target detection is an important task in hyper-spectral image processing. Traditional methods usually impose a stringent assumption on the spectrum distribution of the background and targets, which cannot hold for all the practical situations. This problem can be avoided by sparsity-based method in which each test pixel is represented by a linear combination of a few samples from an overcomplete dictionary. However, classical sparsity model ignores the dictionary structure and cannot guarantee an accurate sparse representation for the test pixel. Motivated by this point, this paper proposes a locality-constrained group sparse representation for target detection. It makes full use of the dictionary structure and preserves the manifold of the original data at the same time, not only ensuring that the correlated training samples belonging to the correct class are used to express the test pixel but also guaranteeing that similar spectrums of HSI pixels will have similar codes. Experimental results on real hyperspectral imagery suggest that the proposed method is more effective than conventional sparsity-based algorithm and the statistics-based methods.