Hyperspectral image denoising based on superpixel segmentation and band segmentation
Existing hyperspectral image denoising algorithms adopt a band-by-band or full-band approach to denoising,which fails to make full use of the similarity of hyperspectral image bands.To address this problem,this paper proposes a hyperspectral image denoising algorithm based on superpixel segmentation and band segmentation.In this paper,we construct a two-layer graph,including the upper and lower layer graphs.First,superpixel segmentation is applied to the hyperspectral image to obtain a series of superpixels.In order to utilize the spatial information on the hyperspectral image and retain the boundary information,the pixels within the superpixels are modeled as nodes with the pixels connected with edges to construct a series of lower layer graphs.In order to utilize the band similarity of the hyperspectral image,superpixel volumes are formed by segmenting along the band dimension based on the superpixel segmentation results with the superpixel volumes modeled as nodes,and the superpixel volumes connected with edges to construct an upper layer graph.Based on the graph structure and graph segmentation,the hyperspectral image denoising problem is reduced to a series of optimization problems,in which the graph Laplacian regularization is redefined using the Kronecker graph product.Finally,experimental results show that the proposed algorithm has a higher mean signal-to-noise ratio,mean structural similarity index measure and erreur relative globale adimensionnelle de synthese compared with the existing algorithms.
hyperspectral image denoisinggraph signal processingsuperpixel segmentationband segmentationgraph Laplacian regularization
万方数据
维普
