Variable bit rate generative compression method based on adversarial learning

Abstract

The invention discloses a variable bit rate generative compression method based on adversarial learning, and belongs to the technical field of communication and digital image processing. The method quantifies the variance of a coding and decoding full convolutional network feature map to train a single generative model to perform variable bit rate compression. The method comprises the following implementation steps: constructing a training and testing data set through image acquisition equipment; constructing a generative compression network based on an auto-encoder structure; alternately training a generative network according to a rate-distortion error calculation unit; calculating a mask threshold value according to the target compression ratio; calculating a mask based on a feature mapchannel redundancy index and a threshold; and performing lossless compression and encoding of masks and feature maps. According to the method, only a single model is trained, but compression resultswith different bit rates can be generated, and the subjective quality and semantic information storage of the reconstructed image have good effects on the limit compression rate below 0.1 bpp.

Description

technical field [0001] The invention provides a data-driven variable bit rate generative compression method, which belongs to the cross technical field of communication and digital image processing. Background technique [0002] Image is an extremely important part of multimedia data. However, limited by communication bandwidth and storage device capacity, original image data needs to be transmitted and stored after specific compression. Relying on the information age, with the rapid development of sensor resolution and Internet scale, the amount of currently available image data is showing an unprecedented explosive growth trend, far exceeding the increase in communication bandwidth and storage capacity. At the same time, the development of existing image compression standards is quite slow, limited by image information entropy and current compression coding methods, image compression technology has obviously entered the bottleneck stage. [0003] The current general image...

AI Technical Summary

Problems solved by technology

This type of method needs to train more models, which brings huge overhead in complexity, time and parameter volume

And changing the quantization method like existing compression standards will cause unpredictable performance loss for learning-based methods

Images