Deep Quality Enhancement of Adaptive Downscaled Coding for Image Compression

An image decoding system, comprising:

• a plurality of block decoders, each associated with pixel blocks coded at a respective size;
• an image reconstruction unit, having inputs coupled respective to outputs of the block decoders;
• and
• a neural network, having an input coupled to an output of the image reconstruction unit.

An image coding method, comprising:

• partitioning a source image into a plurality of pixel blocks;
• selecting, based on content of each pixel block, one of a plurality of coders to code the pixel block, each coder to code pixel blocks at a respective pixel block size;
• coding, at each coder, pixel block(s) input it at the coder's respective pixel block size, the coding including, for coders that operate at pixel block sizes different than the partitioning pixel block sizes, scaling the input pixel block(s) to match the coder's respective pixel block size;
• and
• transmitting coded image data of the pixel blocks.

A coding system, comprising:

• a plurality of block-based coding systems,
• each having inputs for pixel blocks at a common size,
• each coding system comprising a block coder to code pixel blocks at a respective pixel block size,
• at least one image coding system having a downsampler provided at its input for the pixel blocks and an input coupled to its block coder;
• an image segmenter, having an input for a source image, and outputs to each of the image coding systems, the image segmenter to parse the source image into a plurality of pixel blocks, perform content analysis on each pixel block, and route each pixel blocks to one of the coding systems based on the content analysis.

A neural network training system, comprising:

• a plurality of image coding systems,
• each having inputs for pixel blocks at a common size,
• each coding system comprising a block coder and a block decoder adapted to code image at a respective pixel block size,
• at least one image coding system having a downsampler provided at its input for the pixel blocks and an input coupled to its block coder and an upsampler provided at its output for decoded pixel blocks from its block decoder;
• a neural network having an input, an output and an input for network weights;
• a training controller adapted to apply reconstructed pixel blocks from the image coding system to the input of the neural network, to compare data at the output of the neural network to input pixel blocks corresponding to the reconstructed pixel blocks, and to adjust the network weights based on the comparison.