A hyperspectral image compression method based on spectral embedding

Abstract

The application discloses a hyperspectral image compression method based on spectral embedding. The steps are as follows: collecting a static hyperspectral image dataset; extracting a spectral channel index sequence of the image; performing data normalization processing on the spectral channel index sequence to obtain a sequence containing spectral correlation of the image; establishing a hyperspectral image coding and decoding network based on spectral embedding; training the coding and decoding network using the processed sequence data; obtaining spectral embedding corresponding to the image and a trained decoding network; and inputting the spectral embedding of the required image into the decoding network for prediction and inference at the decoding end. The application can provide better reconstruction effect at the same code rate, is significantly superior to existing learning methods, and is comparable to the latest traditional codec. Meanwhile, the hyperspectral image compression method is lightweight, shows faster model convergence in the encoding process, and shows faster decoding in the inference.

Images