Image Compression Using Normalizing Flows

Number: 20210142524
Published: 2021-05-13
Filed: 2020-03-06
Assignee: Disney Enterprises, Inc.
Inventors: Djelouah; Abdelaziz, Helminger; Leonhard Markus, Labrozzi; Scott, Xue; Yuanyi, Doggett; Erika Varis, McPhillen; Jared, Schroers; Christopher Richard
CPC: G06T9/002; H04N19/126; H04N19/60; H04N19/90; H04N19/91
Verdict: Set aside image compression codec, plumbing
Source: Google Patents · FreePatentsOnline

Abstract

According to one implementation, an image compression system includes a computing platform having a hardware processor and a system memory storing a software code. The hardware processor executes the software code to receive an input image, transform the input image to a latent space representation of the input image, and quantize the latent space representation of the input image to produce multiple quantized latents. The hardware processor further executes the software code to encode the quantized latents using a probability density function of the latent space representation of the input image, to generate a bitstream, and convert the bitstream into an output image corresponding to the input image. The probability density function of the latent space representation of the input image is obtained based on a normalizing flow mapping of one of the input image or the latent space representation of the input image.

Background

BACKGROUND

The field of neural image compression has made significant progress with deep learning based approaches. There are two key aspects in neural based image compression. First a mapping from image space to a latent space must be learned. Second, a probability needs to be learned in this new space to allow entropy coding of the latents.

Conventional approaches have addressed the first aspect by proposing neural network architectures to parameterize the encoding and decoding functions typically needed to achieve good compression results. More recently, the primary focus of research has been on the second aspect: trying to accurately model the distribution in the latent space, where there remains a need in the art for additional progress. SUMMARY

There are provided systems and methods for performing image compression using normalizing flows, substantially as shown in and/or described in connection with at least one of the figures, and as set forth more completely in the claims.

Claims

1. An image compression system comprising: a computing platform including a hardware processor and a system memory; a software code stored in the system memory; the hardware processor configured to execute the software code to: receive an input image; transform the input image to a latent space representation of the input image; quantize the latent space representation of the input image to produce a plurality of quantized latents; encode the plurality of quantized latents, using a probability density function of the latent space representation of the input image, to generate a bitstream; and convert the bitstream into an output image corresponding to the input image; wherein the probability density function is obtained based on a normalizing flow mapping of one of the input image or the latent space representation of the input image. || 11. A method for use by an image compression system including a computing platform having a hardware processor and a system memory storing a software code, the method comprising: receiving, by the software code executed by the hardware processor, an input image; transforming, by the software code executed by the hardware processor, the input image to a latent space representation of the input image; quantizing, by the software code executed by the hardware processor, the latent space representation of the input image to produce a plurality of quantized latents; encoding, by the software code executed by the hardware processor, the plurality of quantized latents using a probability density function of the latent space representation of the input image, to generate a bitstream; and converting, by the software code executed by the hardware processor, the bitstream into an output image corresponding to the input image; wherein the probability density function is obtained based on a normalizing flow mapping of one of the input image or the latent space representation of the input image. || 20. An image compression system comprising: a computing platform including a hardware processor and a system memory; a software code stored in the system memory, the software code including a neural network (NN) encoder and an NN decoder; the hardware processor configured to execute the software code to: receive an input image by the NN encoder; transform the input image to a latent space representation of the input image using a normalizing flow mapping of an output of the NN encoder; quantize the latent space representation of the input image to produce a plurality of quantized latents; encode the plurality of quantized latents using a probability density function of the latent space representation of the input image, to generate a bitstream; and convert the bitstream into an output image corresponding to the input image using a normalizing flow mapping of the bitstream and the NN decoder.