Detection of subpixel targets with low fill-fraction: An application to hyperspectral imagery

In hyperspectral imagery, each subpixel target is well known as the target of interest that occupies only a fraction of the pixel area. The remaining part of the pixel is then filled with the background (at the same spatial location). In this paper, we mainly discuss a hyperspectral target detector that is represented as a sparse hyperspectral image (HSI) that ideally contains only subpixel targets with the background suppressed. More precisely, with the help of a pre-learned target dictionary constructed from some online spectral libraries, the given HSI can be decomposed into a sum of low-rank background HSI and a sparse target HSI, where the latter can be directly used as the target detector. However, with this matrix separation model, the detection of the target of interest may fail (or not succeed without a lot of false alarms) when the subpixel target has a very low fill-fraction and especially when its spectra is well matched to the surroundings. To well alleviate this serious real challenge, we prove via some synthetic experiments that learning an additional background dictionary, and when included in the matrix separation model, would be crucial.