The invention discloses a spatial-spectral joint deep convolutional network method for lossy compression of a hyperspectral image, which is used for lossy compression of the hyperspectral image at a low bit rate and comprises the following steps of: integrally inputting the hyperspectral image into a convolutional neural network for training and testing; wherein the encoder and the decoder adopt aconvolution structure to extract spatial-spectral characteristics of a hyperspectral image, an activation function adopted by the network is a generalized split normalization layer, and a bit streamis obtained through a quantizer and entropy encoding to be stored and transmitted; considering the correlation between wavebands, the spectral information is compressed and decompressed by adopting aone-way spectral convolution needle, and then spatial-spectral joint features are extracted through an encoder; wherein the network loss function is a rate-distortion loss function. The loss functionbalances the compression ratio and the image distortion degree, so that the network can learn different compression capacities according to different balance factors, and the performance is remarkablyimproved.

Spatial-spectral joint deep convolutional network method for lossy compression of hyperspectral image

The invention discloses an image compression method based on multi-scale feature coding. The image compression method comprises the steps: obtaining a selection vector through the averaging of the absolute values of gradient spectrums of image features of a training set, and guiding different channel features to select the coding resolution through the selection vector; meanwhile, recovering the features of low-resolution coding at a decoding end through a super-resolution network; and finally recombining the features of low-resolution coding with the features of high-resolution coding to forma complete feature spectrum, and mapping the complete feature spectrum back to an original image. According to the image compression method, difference processing is carried out according to the characteristics of the image characteristics, and the characteristics easy to recover from the context information are transmitted at low resolution, so that the code rate is saved; and for complex fine features, high-resolution transmission is adopted, so that the loss degree is reduced.

Wu Qingbo

Papers

Image compression method based on multi-scale feature coding