Hyperspectral image dimension reduction method and device based on combination of space and spectral features

Abstract

The invention discloses a hyperspectral image dimension reduction method based on combination of space and spectral features, and the method comprises the steps: intercepting a template region of a first frame of image in a hyperspectral image after gray normalization, and determining the gradient probability distribution (GPDF) of each wave band in the template region; obtaining information entropy of hyperspectral spatial correlation intensity and local gradient distribution according to (GPDF); carrying out binary segmentation through the self-adaptive wave band threshold pair to obtain all wave band groups; after weighted fusion and sorting are carried out on each wave band group, the first three wave band groups are connected to serve as template spatial features to be input into a target tracking network; intercepting a target area and a search area from a first frame of the hyperspectral image, respectively obtaining template spectral features and search spectral features, and inputting the template spectral features and the search spectral features into a target tracking network; and sequentially loading all frames of hyperspectral images in the hyperspectral images into the target tracking network to complete hyperspectral image sequence dimensionality reduction.

Images