Data processing method and data processing system

Patent metadata

Field	Value
Publication number	US-12586256-B2
Application number	US-202318373559-A
Country	US
Kind code	B2
Filing date	Sep 27, 2023
Priority date	Dec 28, 2022
Publication date	Mar 24, 2026
Grant date	Mar 24, 2026

Abstract

In a data processing method executed by a data processing system that performs compression and/or decompression of image data, a tensor shape representing compression target data is obtained, compression processing is performed using data having an input shape fixed for each shape of the compression target data as input, and an input shape fixed compressor for outputting compressed data is generated. Then, the data processing system performs compression processing of compression target data using the generated input shape fixed compressor to generate compressed data.

First claim

What is claimed is:

1 . A data processing method executed by a data processing system including a processor that performs compression and/or decompression of image data, the processor executing steps comprising:
obtaining a shape of a tensor representing compression target data; generating an input shape fixed compressor that outputs compressed data by executing compression processing using image data having a fixed input shape for each shape of the compression target data as input; generating the compressed data by executing the compression processing of the compression target data using the generated input shape fixed compressor; executing the compression processing of the compression target data using the input shape fixed compressor when there is an input shape fixed compressor corresponding to the input shape that matches the shape of the obtained compression target data; and executing the compression processing of the compression target data using the general-purpose shape compressor that outputs compression data by executing the compression processing using data having any of the input shapes as input if there is no input shape fixed compressor corresponding to the input shape that matches the shape of the obtained compression target data.

2 . The data processing method according to claim 1 ,
wherein the processor: obtains a shape of a tensor representing image data before execution of the compression process of the decompression target data, generates an output shape fixed decompressor that performs decompression processing for outputting decompression data having an output shape fixed for each shape of the obtained decompression target data, and generates the decompression data by executing the decompression processing of the decompression target data using the generated output shape fixed decompressor.

3 . The data processing method according to claim 2 ,
wherein the processor: executes the decompression processing of the decompression target data using the output shape fixed compressor when there is an output shape fixed compressor corresponding to the output shape that matches the shape of the obtained decompression target data, and executes the decompression processing of the decompression target data using a general-purpose shape decompressor that performs the decompression processing of data having any of the output shapes and outputs the decompression data when there is no output-fixed shape decompressor corresponding to the output shape that matches the shape of the obtained decompression target data.

4 . The data processing method according to claim 3 ,
wherein the processor: counts frequency information on shapes of the compression target data and the decompression target data, based on the frequency information, generates an input shape fixed compressor for each shape of the compression target data and an output shape fixed decompressor for each shape of the decompression target data, and repeats regeneration of the input shape fixed compressor and the output shape fixed decompressor based on the frequency information at a specified trigger.

5 . The data processing method according to claim 4 ,
wherein the shape of the compression target data and the decompression target data includes vertical length and horizontal length of the compression target data and the decompression target data.

6 . The data processing method according to claim 5 ,
wherein the shape of the compression target data and the decompression target data includes color information of the compression target data and the decompression target data.

7 . The data processing method according to claim 5 ,
wherein the processor: collectively executes the compression processing on the compression target data of a number corresponding to the shape of the compression target data, and collectively executes the decompression processing on the decompression target data of a number corresponding to the shape of the decompression target data.

8 . The data processing method according to claim 5 ,
wherein the processor: applies padding to the compression target data, and a plurality of shapes of the compression target data are grouped into input shapes with high frequency in the frequency information, and executes the compression processing of the compression target data and the decompression processing of the decompression target data, using the input shape fixed compressor and the output shape fixed decompressor for each grouped input shape.

9 . The data processing method according to claim 8 ,
wherein the processor: applies padding to the compression target data to match the input shape of the input shape fixed compressor and executes the compression processing of the padded compression target data using the input shape fixed compressor, when there is no input shape fixed compressor of the input shape that matches the shape of the obtained compression target data and when compared that the shape of the compression target data is smaller than the input shape of the input shape fixed compressor within a predetermined range, and executes the decompression processing of the decompression target data using the output shape fixed decompressor and generates the compression target data before compression by removing the padding from the decompressed data, when there is no output shape fixed decompressor of the output shape that matches the shape of the obtained decompression target data and when compared that the shape of the decompression target data is smaller than the output shape of the output shape fixed decompressor within a predetermined range.

10 . The data processing method according to claim 9 ,
wherein the compressed data holds information on the padding applied to the compression target data.

11 . The data processing method according to claim 2 ,
wherein the processor: collectively executes the compression processing on a specified number of the compression target data for each shape of the compression target data, and collectively executes the decompression processing on a specified number of the decompression target data for each shape of the decompression target data.

12 . The data processing method according to claim 2 ,
wherein each of the input shape fixed compressor and the output shape fixed decompressor is a learning-based compressor and a learning-based decompressor generated by training a neural network.

13 . A data processing system, comprising:
a processor for executing compression and/or decompression of image data, wherein the processor: obtains a shape of a tensor representing compression target data, generates an input shape fixed compressor that outputs compressed data by executing compression processing using image data having a fixed input shape for each shape of the compression target data as input, and generates the compressed data by executing the compression processing of the compression target data using the generated input shape fixed compressor, executes the compression processing of the compression target data using the input shape fixed compressor when there is an input shape fixed compressor corresponding to the input shape that matches the shape of the obtained compression target data, and executes the compression processing of the compression target data using the general-purpose shape compressor that outputs compression data by executing the compression processing using data having any of the input shapes as input if there is no input shape fixed compressor corresponding to the input shape that matches the shape of the obtained compression target data.

Assignees

• Hitachi Ltd
• Hitachi Vantara Ltd

Inventors

• Suzuki Akifumi
• Akutsu Hiroaki

Classifications

• G06T9/002

  using neural networks · CPC title

• G06T9/20 Primary

  Contour coding, e.g. using detection of edges · CPC title