Hyperspectral Anomaly Detection by Data Sphering and Sparsity Density Peaks

Background-Annihilated Target-Constrained Interference-Minimized Filter (TCIMF) for Hyperspectral Target Detection

“…Fortunately, in a recent work [30], Chang et al interpreted and re-derived GoDec as an OSP-based GoDec (OSP-GoDec) which only required determining the parameter k. In particular, this OSP-GoDec took advantage of VD in [1,[32][33][34][35][36] and the minimax singular value decomposition (MX-SVD) developed in [37] to automatically determine the value of k. This significant benefit allows GoDec to be implemented without determining all the four parameters, r, k, ε, γ empirically. So, it makes GoDec more accessible and increase its utilization in many future works to come [14][15]. The reason that GoDec was used in this is simply because GoDec is widely used in the literature and the details along with its implementations are available on the website 1 .…”

“…onto Ω via the projection P Ω . For various comparative analyses can be found in references [14][15]17,30,38], specifically, reference [38]. Those who are interested in Scare encouraged to consult [17,38] for more details of SC interpretations and derivations In particular, Chang in [38] provided how the concept of SC was derived and also offered different versions of SC.…”

“…Recently, LRaSMD was developed to represent a data space X as a three-component subspace decomposition, = + + X L S N with the low rank space L and sparse space S corresponding to the BKG space and target space respectively and the noise space is specified by N. By taking advantage of this LRaSMD model, we can make use of OSP to annihilate the BKG prior to TCIMF via P ⊥ L X in (15). By means of the low rank space L generated by LRaSMD model we can find additional unwanted BKG signals from L for further improvement on BKG suppression.…”

“…In this case, data sphering (DS) [2,chapter 6] is then applied to annihilate BKG prior to target detection. This DS was further used to expand OSP's capability in improving target detection [14] and anomaly detection [15]. A second one is to represent a data space X as a decomposition of a low rank subspace L and sparse subspace S so that L and S can be used to specify the BKG and target spaces, respectively.…”

Matrix Autoregressive Model for Hyperspectral Anomaly Detection

“…Considering the noise is also sparse, some works attempt to separate the noise from the sparse space and to find the true anomalies, which is formulated in the classical GoDec method [23]. Similar to this idea, a component decomposition analysis method is proposed to treat the HSI as a combination of the principal components, independent components and a noise component [24].…”