Hyperspectral remote sensing image compression method, device and product based on hybrid hyper-prior

Abstract

The invention discloses a hyperspectral remote sensing image compression method, device and product based on hybrid hyper-prior, and the method comprises the steps: carrying out the conversion between an original image and compact potential representation through employing a CNN and MLP-based backbone analysis network and a synthesis network, and obtaining a better potential representation feature. Besides, a space-pass-to-attention module is used, the representation capability of a CNN and MLP hybrid network module is further enhanced, and redundant information in the hyperspectral image is deeply mined. When the entropy model is constructed, super-prior mining of local redundant information based on CNN is used, super-prior extraction of non-local redundant information based on Transform is used, the two are fused, parameters of the entropy model are predicted, the accuracy of entropy model construction is improved, and the final image compression performance is improved.

Images