Apparatus and method for compressing image for machine vision

Patent metadata

Field	Value
Publication number	US-12530781-B2
Application number	US-202217697463-A
Country	US
Kind code	B2
Filing date	Mar 17, 2022
Priority date	Mar 18, 2021
Publication date	Jan 20, 2026
Grant date	Jan 20, 2026

Abstract

Disclosed herein is a method for compressing an image for machine vision, the method including detecting objects in an input image using an object detection network, generating a foreground image including bounding boxes corresponding to the objects and a background image, which is an image acquired by excluding the bounding boxes from the input image, encoding the foreground image and the background image, and decoding the encoded foreground image and the encoded background image.

First claim

What is claimed is:

1 . A method for compressing an image for machine vision, comprising:
detecting objects in an input image using an object detection network; generating a foreground image, including bounding boxes corresponding to the objects, from the input image; encoding the foreground image; and wherein generating the foreground image includes:
determining initial bounding boxes corresponding to the objects,
extending sizes of the initial bounding boxes, and
generating the foreground image including the bounding boxes with extended sizes,
wherein in response to a ratio between a height and a width of a bounding box being greater than a preset first ratio or less than a reciprocal of the preset first ratio, the height and the width of the bounding box are extended by an average value of the height and the width, and wherein in response to the ratio between the height and the width of the bounding box being equal to or less than the preset first ratio or being equal to or greater than the reciprocal of the preset first ratio, the height and the width of the bounding box are extended by a smaller one of the height and the width.

2 . The method of claim 1 , wherein the method further comprises:
determining on a scaling factor for the foreground image.

3 . The method of claim 2 , wherein the method further comprises generating a background image corresponding to a remaining image excluding the foreground image from the input image, and
wherein a scaling factor for the background image is equal to or less than & the scaling factor for the foreground image.

4 . The method of claim 3 , wherein
the foreground image is encoded by using a first quantization parameter (QP), and wherein the background image is encoded by using a second quantization parameter, which is greater than the first quantization parameter.

5 . The method of claim 2 , wherein the scaling factor is obtained by modifying an initial scaling factor.

6 . The method of claim 1 , wherein:
the input image corresponds to a thermal infrared image, and the object detection network corresponds to a network adjusted using training data including thermal infrared images and RGB images.

7 . An apparatus for compressing an image for machine vision, comprising:
an object detector configured to detect objects in an input image using an object detection network; an image generator configured to:
determine initial bounding boxes corresponding to the objects,
extend sizes of the initial bounding boxes by, and
generate a foreground image, including the bounding boxes with extended sizes; and
an encoder configured to encode the foreground image, wherein in response to a ratio between a height and a width of a bounding box being greater than a preset first ratio or less than a reciprocal of the preset first ratio, the height and the width of the bounding box are extended by an average value of the height and the width, and wherein in response to the ratio between the height and the width of the bounding box being equal to or less than the preset first ratio or being equal to or greater than the reciprocal of the preset first ratio, the height and the width of the bounding box are extended by a smaller one of the height and the width.

8 . The apparatus of claim 7 , wherein:
the apparatus further includes: a downsampler configured to: determine a scaling factor for the foreground image.

9 . The apparatus of claim 8 , wherein the image generator is further configured to generate a background image corresponding to a remaining image excluding the foreground image from the input image,
wherein the downsampler is further configured to determine a scaling factor for the background image and wherein the scaling factor for the background image is equal to or less than the scaling factor for the foreground image.

10 . The apparatus of claim 9 , wherein:
the encoder comprises a first encoder to encode the foreground image and a second encoder to encode the background image, wherein: the first encoder is configured to encode the foreground image using a first quantization parameter (QP), and the second encoder is configured to encode the background image using a second quantization parameter, which is greater than the first quantization parameter.

11 . The apparatus of claim 8 , wherein the scaling factor is obtained by modifying an initial scaling factor.

12 . The apparatus of claim 7 , wherein:
the input image corresponds to a thermal infrared image, and the object detection network corresponds to a network adjusted using training data including thermal infrared images and RGB images.

Assignees

• Electronics & Telecommunications Res Inst
• Univ Konkuk Ind Coop Corp

Inventors

• Yoon Kyoung-Ro
• Lim Han-Shin
• Choo Hyon-Gon
• Kim Shin
• Lee Ye-Gi

Classifications

• G06T9/002

  using neural networks · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/10048

  Infrared image · CPC title

• G06T2207/10024

  Color image · CPC title