Systems and methods for neural network based data compression

A system for data compression, comprising:

• a computing device comprising at least a memory and a processor;
• a preprocessing subsystem comprising a first plurality of programming instructions that, when operating on the processor, cause the computing device to perform one or more preprocessing operations on an uncompressed dataset;
• a discrete cosine transform (DCT) subsystem comprising a second plurality of programming instructions that, when operating on the processor, cause the computing device to:
• perform a DCT operation on the uncompressed dataset, and create a plurality of subbands for the uncompressed dataset, wherein each subband represents a frequency range;
• and
• divide the plurality of subbands into multiple groups, wherein the multiple groups comprise a first low frequency group, a second low frequency group, and a high frequency group;
• a compression subsystem comprising a third plurality of programming instructions that, when operating on the processor, cause the computing device to implement a latent feature encoding network, wherein the latent feature encoding network comprises:
• a first kernel, the first kernel configured to provide output to a first plurality of residual blocks, the first plurality of residual blocks configured to provide output to a first attention network, the first attention network configured to provide output to a second plurality of residual blocks, the second plurality of residual blocks configured to provide output to a second kernel, and the second kernel configured to provide output to a second attention network, and wherein the latent feature encoding network is configured and disposed to generate a latent space representation corresponding to the uncompressed dataset;
• a first arithmetic coding subsystem comprising a fourth plurality of programming instructions that, when operating on the processor, cause the computing device to perform compression on one or more of the plurality of subbands, thereby creating a first compressed bitstream;
• a hyperprior latent feature summarization module configured and disposed to receive input from the second attention network, and provide a refined and disentangled representation of one or more latent features as output;
• and
• a second arithmetic coding subsystem comprising a fifth plurality of programming instructions that, when operating on the processor, cause the computing device to perform compression on the output of the hyperprior latent feature summarization module, thereby creating a second compressed bitstream;
• wherein the first compressed bitstream and second compressed bitstream together represent a compressed version of the uncompressed dataset.