An Image Processing Method Based on Matrix Variable Variational Autoencoder

Abstract

The invention discloses an image processing method based on a matrix variable variational autoencoder, which can solve the problem of image vectorization processing destroying the spatial structure, and further facilitates image reconstruction, denoising and completion. Different from the traditional VAE, this method uses the inherent representation of the image-2D matrix to describe the input of the model, hidden layer features, latent variable distribution characteristic parameters, etc., and derives the target of the new model by using the definition and related properties of the matrix Gaussian distribution The explicit expression of the function, and then use the stochastic gradient descent algorithm to solve the model parameters. In this model, since the modeling process involved in the present invention is oriented to matrix variables, it can better model the spatial structure and statistical information of image data, thereby improving the quality of image reconstruction, better removing noise and image Completion.

Description

technical field [0001] The invention belongs to the field of computer vision and machine learning, and in particular relates to an image processing method based on a matrix variable variational autoencoder. Background technique [0002] Image reconstruction, denoising and completion are important contents of image processing. Variational Autoencoder (VAE) is widely used in image processing related fields because it can well model the probability distribution of image data. VAE usually consists of an inference model (encoder) and a generation model (decoder). The objective function of the model includes two items: one is the reconstruction error of the image, which is generally measured by mean square error or cross entropy; the other is The Kullback-Leibler (KL) divergence is used to measure the similarity between the posterior of the feature latent variables learned based on the inference model and the prior distribution of the feature assumptions, which is equivalent to a ...

AI Technical Summary

Problems solved by technology

[0004] The technical problem to be solved by the present invention is to provide a method for image processing based on a matrix variable variational autoencoder (Matrix-variate Variational Autoencoder, MVVAE), which can solve the problem of image vectorization processing destroying the spatial structure, thereby benefiting the image Refactoring, denoising and completion

Images