Unsupervised Segmentation Of Hyper-Spectral Images Based On Superpixels

Hyperspectral imaging technology is an imaging technology rise in last century 80's,it can be imaged objects in visible light and near infrared wavelength range.Based on the advantages of spectral resolution and imaging range,it can detect a lot of information which can not be detected by ordinary imaging equipment,and has a high application value in economic construction and military affairs.Hyperspectral image segmentation is the basis for the analysis and understanding of hyperspectral images,it is important to study it.Superpixel is an image segmentation technology refers to the adjacent pixels with irregular texture,color,brightness and other similar features which have a certain visual meaning of pixel blocks.It makes use of the similarity of the features between pixels to group the pixels,and uses a small number of superpixels instead of a large number of pixels to express the image features,greatly reduce the complexity of the image postprocessing.In this paper,we propose an unsupervised segmentation algorithm for hyperspectral images based on super pixels,which has high superiority in edge preserving and automation.Hyperspectral images have a large amount of information and high data-to-spectrum correlation.Therefore,we also studies the method of reducing the hyperspectral image in this peper.In this paper,the unsupervised segmentation algorithm of hyperspectral image is studied.Firstly,the PCA and MNF dimension reduction methods of hyperspectral image are introduced and compared.Then,we introduce three kinds of commonly used superpixel segmentation algorithms:SLIC superpixel,Turbopixel superpixel and watershed superpixel segmentation algorithm.Through the contrast experiment,we choose the SLIC superpixel segmentation method to carry on our experiment.In the end,the image segmentation is performed by using the coefficient of similarity as the similarity measure and the adjacent pixels are merged.By comparing with the K-means hyperspectral segmentation algorithm and the gradient reconstruction watershed segmentation algorithm,the superiority of the algorithm in the edge preservation and automation level is demonstrated.