Image coding, decoding and compression method based on depth Gaussian process regression

Abstract

The invention discloses an image coding method based on deep Gaussian process regression, and the method comprises the steps: obtaining a bottleneck layer multi-channel feature of a to-be-coded image through employing a coding convolutional neural network, and taking the feature as a first feature map; quantifying each feature in the first feature map into an integer to obtain a second feature map; on the basis of an autoregression model and a hyper-prior model of deep Gaussian process regression, weighting and combining a plurality of Gaussian mixture distribution coding features of Gaussian distribution for each channel of the second feature map, and generating a feature binary code stream; super-prior information obtained by the super-prior model is coded into a super-prior binary code stream; and combining the super-prior binary code stream and the feature binary code stream to obtain a binary code stream of the compressed image. A non-parametric depth Gaussian process regression method is adopted for autoregression modeling, posterior distribution output by depth Gaussian process regression serves as a mean value of a Gaussian mixture model, the uncertainty of mean value estimation can be flexibly obtained, and therefore more accurate mean value estimation is obtained.

Description

technical field [0001] The present invention relates to the field of image compression, in particular, to an image encoding, decoding and compression method based on deep Gaussian process regression. Background technique [0002] In recent years, deep neural networks have been widely used in the field of image compression. In the end-to-end image compression methods emerging in recent years, deep neural networks replace the transformation, quantization and entropy coding modules in traditional image compression methods. The current mainstream end-to-end image compression method adopts the architecture of variational autoencoder. In this architecture, the accuracy of modeling the feature entropy model of the bottleneck layer has attracted widespread attention because it greatly affects the size of the bit rate. [0003] Early entropy models assumed that the elements of bottleneck layer features are independent of each other. After a literature search of the prior art, it w...

AI Technical Summary

Problems solved by technology

Therefore, when the estimate of the mean is not accurate, a larger residual error will be generated, resulting in a smaller corresponding probability value and an increased code rate

At present, the mixed Gaussian distribution is based on a linear combination of a certain number of Gaussian distribution means to estimate the mean value, rather than estimating the distribution of the mean value, so it is more susceptible to the influence of the input image, so the estimated value of the bottleneck layer features will be inaccurate and produce Deviation, so that the estimated probability used for encoding is far away from the peak of the Gaussian distribution, resulting in an increase in the bit rate of the compressed image, which deteriorates the rate-distortion performance

Images