Method, apparatus, and storage medium for encoding/decoding feature map for machine vision

Patent metadata

Field	Value
Publication number	US-12556743-B2
Application number	US-202318302719-A
Country	US
Kind code	B2
Filing date	Apr 18, 2023
Priority date	Apr 18, 2022
Publication date	Feb 17, 2026
Grant date	Feb 17, 2026

Abstract

Disclosed herein are a method, an apparatus, and a storage medium for encoding/decoding a feature map for machine vision. In an embodiment, an encoding method includes generating a multi-channel feature map by aligning multiple feature maps, generating a converted feature map by performing conversion on the multi-channel feature map and performing encoding on the converted feature map.

First claim

What is claimed is:

1 . An encoding method, comprising:
generating a multi-channel feature map by aligning multiple feature maps; generating a converted feature map by performing conversion on the multi-channel feature map; and performing encoding on the converted feature map, wherein performing the encoding comprises: rearranging channels of the converted feature map based on a correlation between the channels of the converted feature map.

2 . The encoding method of claim 1 , wherein generating the multi-channel feature map comprises:
scaling feature maps having different resolutions to an identical resolution.

3 . The encoding method of claim 1 , wherein generating the converted feature map comprises:
extracting representative values for respective channels of the multi-channel feature map.

4 . The encoding method of claim 3 , wherein the representative values for respective channels are adjusted through a fully-connected layer.

5 . The encoding method of claim 3 , wherein the multi-channel feature map is adjusted based on the representative values for respective channels.

6 . The encoding method of claim 1 , wherein generating the converted feature map is performed based on a conversion network including one or more of a pooling layer, a fully-connected layer, a convolution layer, and an activation function layer.

7 . An encoding method, comprising:
generating a multi-channel feature map by aligning multiple feature maps; generating a converted feature map by performing conversion on the multi-channel feature map; and performing encoding on the converted feature map, wherein: the converted feature map includes a number of channels different from that of the multi-channel feature map, and generating the converted feature map comprises adjusting a quantization parameter based on a change in the number of channels in the converted feature map.

8 . The encoding method of claim 1 , wherein generating the converted feature map comprises:
calculating importance levels for respective channels of the multi-channel feature map and converting a number of channels based on the importance levels for the respective channels.

9 . The encoding method of claim 1 , wherein the correlation between the channels is calculated based on an average value for each channel, a median value for each channel, or a Mean Squared Error (MSE) between the channels.

10 . The encoding method of claim 1 , wherein performing the encoding comprises:
converting the converted feature map into a feature map of one frame.

11 . The encoding method of claim 10 , wherein performing the encoding further comprises:
performing tile-based encoding on the converted feature map of one frame.

12 . A decoding method, comprising:
performing size reconstruction on a converted feature map; and generating a reconstructed feature map by reconstructing a number of channels in the converted feature map, wherein the reconstructed feature map includes feature maps having different resolutions or different numbers of channels, and wherein the channels of the converted feature map are rearranged based on a correlation between the channels of the converted feature map.

13 . The decoding method of claim 12 , wherein performing the size reconstruction comprises:
differently configuring layers for reconstruction depending on a resolution of a feature map desired to be reconstructed.

14 . The decoding method of claim 12 , wherein the reconstructed feature map is generated based on a reconstruction network including one or more of a pooling layer, a fully-connected layer, a convolution layer, and an activation function layer.

15 . The decoding method of claim 12 , wherein performing the size reconstruction comprises:
performing size reconstruction using size information of an original feature map.

16 . The decoding method of claim 12 , wherein the size reconstruction or the reconstruction of the number of channels is performed using residual-based reconstruction.

Assignees

• Electronics & Telecommunications Res Inst
• Iucf Korea Aerospace Univ
• Industry Univ Cooperation Foundation Korea Univ

Inventors

• Lee Joo-Young
• Kim Youn-Hee
• Jeong Se-Yoon
• Choi Jin-Soo
• Kim Dong-Ha
• Kim Jae-Gon
• Yoon Yong-Uk

Classifications

• H04N19/30

  using hierarchical techniques, e.g. scalability (H04N19/63 takes precedence) · CPC title

• H04N19/186

  the unit being a colour or a chrominance component · CPC title

• H04N19/103

  Selection of coding mode or of prediction mode · CPC title

• H04N19/174

  the region being a slice, e.g. a line of blocks or a group of blocks · CPC title

• H04N19/172

  the region being a picture, frame or field · CPC title