Geometric separation and applications to hyperspectral image analysis
Monday, October 22, 2018 - 3:05pm - 3:55pm
Combining different basis to build a more efficient combined representation of complex data has a long history in harmonic analysis and signal processing. More recently, a rigorous mathematical framework was proposed to formalize the problem of separating data into morphologically distinct components based on the microlocal properties of the data. According to this framework, we model an hyperspectral image as a mixture of two distinct morphological components - a texture and cartoon-like components - and we apply a combined dictionary including a local cosine and a multiscale shearlet basis. Taking advantage of the property that each basis sparsely represents only one data component and that the two basis are highly incoherent, we are able to geometrically separate data so that they are better suited for the application of sparse representation based classifiers. Numerical results using real-world hyperspectral images show the efficacy of the proposed framework for classification under a variety of adverse conditions.
42C15 · 42C40 · 65T60