System and method for video processing based on quantization parameter

Patent metadata

Field	Value
Publication number	US-2017264902-A1
Application number	US-201615065248-A
Country	US
Kind code	A1
Filing date	Mar 9, 2016
Priority date	Mar 9, 2016
Publication date	Sep 14, 2017
Grant date	—

Abstract

Various aspects of a system and method to process video based on quantization parameters are disclosed herein. In an embodiment, the method includes extraction of a plurality of features to capture texture information of an image block. A neural network regressor is trained to map the extracted plurality of features to determine an optimal quantization parameter. The image block is encoded by use of the determined optimal quantization parameter.

First claim

What is claimed is:

1 . A system for video processing, said system comprising:
one or more circuits in a video processor, said one or more circuits being configured to:
extract a plurality of features to capture texture information of an image block:
train a neural network regressor to map said extracted plurality of features to determine an optimal quantization parameter; and
encode said image block by use of said determined optimal quantization parameter.

2 . The system according to claim 1 , wherein said one or more circuits are configured to encode said image block of a first image frame with a plurality of quantization parameters to generate a plurality of reconstructed image blocks for said image block of said first image frame.

3 . The system according to claim 2 , wherein said one or more circuits are configured to utilize an image quality measure for each of said plurality of reconstructed image blocks for said determination of said optimal quantization parameter for said image block.

4 . The system according to claim 3 , wherein said image quality measure is a full-reference image quality measure based on convolution neural network.

5 . The system according to claim 3 , wherein said one or more circuits are configured to generate a score for each of said plurality of reconstructed image blocks by use of said image quality measure, wherein said score indicates a measure of visual quality of each of said plurality of reconstructed image blocks.

6 . The system according to claim 5 , wherein a value of said determined optimal quantization parameter is highest amongst values of said plurality of quantization parameters, and wherein said value of said determined optimal quantization parameter is greater than or equal to a pre-specified image quality threshold.

7 . The system according to claim 1 , therein said one or more circuits are configured to generate a training data set based on said extracted plurality of features and corresponding said determined optimal quantization parameter of said image block.

8 . The system according to claim 7 , wherein said generated training data set includes a plurality of features of other image blocks of a first image frame of video content and corresponding optimal quantization parameters, wherein said plurality of features of other image blocks are extracted to capture texture information of said other image blocks.

9 . The system according to claim 7 , wherein said one or more circuits are configured to utilize said generated training data set for said training of said neural network regressor.

10 . The system according to claim 1 , said one or more circuits are configured to determine a mapping function between said extracted plurality of features and said determined optimal quantization parameter of said image block based on said trained neural network regressor.

11 . The system according to claim 1 , wherein said neural network regressor is a feed-forward neural network based regression model.

12 . The system according to claim 1 , wherein said one or more circuits are configured to predict another optimal quantization parameter for another image block of a second image frame based on said trained neural network regressor.

13 . A method for video processing, said method comprising:
extracting, by one or more circuits in a video processor, a plurality of features to capture texture information of an image block; training, by said one or more circuits, a neural network regressor to map said extracted plurality of features to determine an optimal quantization parameter; and encoding, by said one or more circuits, said image block by use of said determined optimal quantization parameter.

14 . The method according to claim 13 , further comprising encoding, by said one or more circuits, said image block of a first image frame with a plurality of quantization parameters to generate a plurality of reconstructed image blocks for said image block of said first image frame.

15 . The method according to claim 14 , further comprising utilizing, by said one or more circuits, an image quality measure for each of said plurality of reconstructed image blocks for said determination of said optimal quantization parameter for said image block.

16 . The method according to claim 15 , further comprising generating, by said one or more circuits, a score for each of said plurality of reconstructed image blocks by use of said image quality measure, wherein said score indicates a measure of visual quality of each of said plurality of reconstructed image blocks.

17 . The method according to claim 13 , further comprising generating, by said one or more circuits, a training data set based on said extracted plurality of features and corresponding said determined optimal quantization parameter of said image block, wherein said generated training data set is used for said training of said neural network regressor.

18 . The method according to claim 13 , further comprising determining, by said one or more circuits, a mapping function between said extracted plurality of features and said determined optimal quantization parameter of said image block based on said trained neural network regressor.

19 . The method according to claim 13 , wherein said neural network regressor is a feed-forward neural network based regression model.

20 . The method according to claim 13 , further comprising predicting, by said one or more circuits, another optimal quantization parameter for another image block of a second image frame based on said trained neural network regressor.

Assignees

• Sony Corp

Inventors

• Ye Peng
• Xu Xun

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/048

  Activation functions · CPC title

• H04N19/124 Primary

  Quantisation · CPC title

• H04N19/154

  Measured or subjectively estimated visual quality after decoding, e.g. measurement of distortion (use of rate-distortion criteria H04N19/147) · CPC title

• H04N19/176

  the region being a block, e.g. a macroblock · CPC title