Small target detection in infrared image via sparse representation

A novel method of small target detection in infrared image via sparse representation is proposed in this paper. At the beginning, plenty small target image patches are generated based on mathematical model to construct an overcomplete dictionary which can be used to approximate a test image patch. However, all image patches distribute only in a small part of the whole space because their entries are non-negative which denote image intensity. To make them easier to discriminate, each image patch is changed into a vector by stacking its columns and zero-mean-unit-norm normalization is applied. Then patch at each pixel position of test image is approximated linearly with fewest dictionary atoms and positive correlation index(PCI) is calculated based on the coefficient to evaluate the similarity between a test image patch and small target dictionary. Finally, a two-dimensional map will be produced based on PCI after lexicographically scanning and target can be located with a suitable threshold. Experimental results of infrared images with a wide variety of background demonstrate the effectiveness and robustness of the proposed method.