Abstract
Existing sparsity-based hyperspectral image (HSI) target detection methods have two key problems. 1) The background dictionary is locally constructed by the pixels between the inner and outer windows, surrounding and enclosing the central test pixel. The dual-window strategy is intricate and might result in impure background dictionary deteriorating the detection performance. 2) For an unbalanced binary classification problem, the target dictionary atoms are generally inadequate compared with the background dictionary, which might yield unstable performance. For the issues, this article proposes a novel structurally incoherent background and target dictionaries (SIBTD) learning model for HSI target detection. Specifically, with the concept that the observed HSI data is composed of low-rank background, sparsely distributed targets, and dense Gaussian noise, the background and target dictionaries can be jointly derived from the observed HSI data. Additionally, the introduction of structural incoherence can enhances the discrimination between the target and background dictionaries. Thus, the developed model can not only lead to a pure and unified background dictionary but also augment the target dictionary for improved detection performance. Besides, an efficient optimization algorithm is devised to solve SIBTD model, and the performance of SIBTD is verified on three benchmark HSI datasets in comparison with several state-of-the-art detectors.
Highlights
U NLIKE the conventional RGB image, which is composed of only three spectral bands, hyperspectral imaging sensors can sense reflected light in hundreds or even more than a thousand bands, and a 3-D data structure called hyperspectral image (HSI) can be obtained with two spatial dimensions and one spectral dimension [1], [2]
In sparse representation-based target detector (SRD), a test pixel in an HSI is first represented as a sparse linear combination of a number of samples from the target and background union dictionary, and the detection decision is made by checking which dictionary can yield the smallest representation residual [15]
Once the optimal solutions are obtained, the final augmented and discrimination enhanced target dictionary is gotten as DT = [DTC, S], which can be utilized as the universal target dictionary for target detection by combining with background dictionary and different representation strategies
Summary
U NLIKE the conventional RGB image, which is composed of only three spectral bands, hyperspectral imaging sensors can sense reflected light in hundreds or even more than a thousand bands, and a 3-D data structure called hyperspectral image (HSI) can be obtained with two spatial dimensions and one spectral dimension [1], [2]. Based on the binary hypothesis, the sparse representation-based binary hypothesis detector (SRBBHD) has been suggested wherein the test pixel is modeled by the background dictionary under the null hypothesis or by the background and target union dictionary under the alternative hypothesis [18]. By considering the data characteristics and prior knowledge of HSI, the background dictionary should have the property of low rankness, while the target components are sparsely distributed in spatial. This article proposes to directly learn background and target dictionaries from the observed HSI data via sparsity and low-rank constraints, and a novel structurally incoherent background and target dictionaries (SIBTD) learning model is developed for HSI target detection. 1) SIBTD can learn target and background dictionaries from observed HSI data via low-rank and sparsity constraints.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
