Method for Image Compressed Sensing based on Deep Learning via Learnable Spatial-Spectral transformation

摘要

We propose a deep learning image compression and restoration system based on a learnable spatial-spectral transformation technique. The deep learning image compression and decompression system based on the learnable spatial-spectral transformation technique proposed in the present invention passes through a transformation process that includes a CNN structure between the low-pass filter and the high-pass filter of the wavelet, so that the input image passes through a plurality of spatial-spectral An image conversion unit that generates a real subband image, an extrusion sampling unit that performs compression sampling in a block compression sensing method through a fully connected layer, and a block compression measurement that generates a block compression measurement value from a plurality of spatial-spectral subband images An initial reconstruction unit that generates an initially reconstructed spatial-spectral subband through an initial reconstruction process through a fully connected layer using block compression measurement values for sub- and multiple spatial-spectral subband images, an hourglass-shaped depth A deep reconstruction unit that creates a refined deep reconstruction space-spectral subband while expanding and compressing the number of channels through the reconstruction process, and a deep reconstruction space-spectral subband of an image through inverse transformation of a learnable space-spectral transformation It includes an image inverse transform unit that finally restores the form.