Hyperspectral image segmentation based on the estimation of the multiway signal subspace dimension (Conference Presentation)

Abstract

One of the objectives of image processing is to detect the region of interest (ROI) in the given application, and then perform characterization and classification of these regions. In HyperSpectral Images (HSI) the detection of targets in an image is of great interest for several applications. However, edge detection is not considered by the community. Moreover, it is less easy to define what can be a contour in a HSI than in a conventional image where the edges are defined as a rapid change in the intensity of the pixels. In fact, because the spectral data (hyperspectral pixels) consist of several hundreds of values, it becomes more difficult to define the variations between two pixels. One can found in the literature a few edge detection works on multi or hyperspectral images, based on statistical similarity criteria (maximum likelihood, Mahalanobis distance), or based on a geometric measurement approach (study of distances or angles between pixel vectors), some works also adapt to multispectral data some classical convolution filters (Sobel gradient, etc.), or mathematical morphology methods. But all these methods are more suited to the multispectral than the hyperspectral data because they are put in difficulty by the large amount of information known as Hugues phenomenon. In this paper we propose a new multidimensional method, based on tensor modelling of HSI, and using a local approach thanks to subtensors rank estimation to perform edge detection. This could be a good way to select regions of interest which are known to be useful when small target detection is needed. Generally, when ROI containing targets is previously selected, the detection results are better. The ROI selection for large HSI containing objects with large and small sizes, is based on tensorial modelling, and an estimation of local signal subspace rank variations. Examples of the results obtained by the proposed method show that the local variations of signal subspace rank for subtensors of 5x5 pixels on 143 bands centered on each pixel of the image. The processed real-world HSI containing an almost uniform background and 4 rows of different sizes of panels. The obtained results concerning the evolution of the local rank of subtensors centered on each pixel of the HSI, show that the local rank of the signal subspace and the panels are linked. According to these promising and convincing results, in this paper we propose an algorithm for edge detection based on local rank variations.