Image segmentation method, device and medium

Patent metadata

Field	Value
Publication number	US-12400418-B2
Application number	US-202218148812-A
Country	US
Kind code	B2
Filing date	Dec 30, 2022
Priority date	Feb 28, 2022
Publication date	Aug 26, 2025
Grant date	Aug 26, 2025

Abstract

An image segmentation method includes: obtaining an image to be segmented containing a target object; performing at least one image feature fusion processing on associated feature points based on the image to be segmented, and extracting global feature information during each image feature fusion processing, wherein the associated feature points are at least two feature points having a location association relation; and determining, based on the global feature information extracted, a segmentation mask for the target object.

First claim

What is claimed is:

1. An image segmentation method, comprising:
obtaining an image to be segmented containing a target object;
performing at least one image feature fusion processing on associated feature points based on the image to be segmented, and extracting global feature information during each image feature fusion processing, wherein the associated feature points are at least two feature points having a location association relation; and
determining, based on the global feature information extracted, a segmentation mask for the target object;
wherein the image feature fusion processing comprises first fusion processing and second fusion processing, and the global feature information comprises at least one first global feature map and at least one second global feature map; and
wherein performing the at least one image feature fusion processing on associated feature points based on the image to be segmented and extracting the global feature information during each image feature fusion processing, comprises:
extracting initial image features of the image to be segmented;
obtaining the at least one first global feature map by performing at least one first fusion processing on associated feature points based on the initial image features, and by extracting global feature information during each first fusion processing, wherein different first global feature maps have different resolutions; and
obtaining the at least one second global feature map by performing at least one second fusion processing on associated feature points based on a first global feature map with a lowest resolution, and by extracting global feature information during each second fusion processing, wherein different second global feature maps have different resolutions;
wherein determining, based on the global feature information extracted, the segmentation mask for the target object comprises:
obtaining a third global feature map by performing up-sampling processing on a second global feature map with a highest resolution and assigning image features of feature points in the second global feature map with the highest resolution to new adjacent feature points, wherein the third global feature map has the same resolution as that of the image to be segmented; and
obtaining the segmentation mask for the target object by performing linear mapping processing on the third global feature map.

2. The method of claim 1 , wherein obtaining the at least one first global feature map by performing the at least one first fusion processing on associated feature points based on the initial image features, and by extracting the global feature information during each first fusion processing, comprises:
obtaining, for each first fusion processing, a first feature map corresponding to the first fusion processing, wherein a feature map corresponding to the initial image features is a first feature map corresponding to 1 st first fusion processing;
obtaining a second feature map by performing a shift operation on feature points in the first feature map; and
obtaining a first global feature map by performing the first fusion processing on associated feature points in the first feature map and associated feature points in the second feature map, and by extracting the global feature information during the first fusion processing.

3. The method of claim 2 , wherein obtaining the first feature map corresponding to the first fusion processing, comprises:
for N th first fusion processing, obtaining a first global feature map obtained in previous first fusion processing, wherein N is a positive integer greater than 1; and
obtaining a first feature map corresponding to the N th first fusion processing by performing down-sampling processing on the first global feature map obtained, and by splicing image features of reduced feature points into image features of adjacent feature points.

4. The method of claim 1 , wherein obtaining the at least one second global feature map by performing the at least one second fusion processing on associated feature points based on the first global feature map with the lowest resolution, and by extracting the global feature information during each second fusion processing, comprises:
obtaining a second global feature map with a lowest resolution by performing at least one third fusion processing on associated feature points based on the first global feature map with the lowest resolution, wherein the first global feature map with the lowest resolution has the same resolution as that of the second global feature map with the lowest resolution; and
obtaining the at least one second global feature map by performing at least one second fusion processing on associated feature points based on the second global feature map with the lowest resolution, and by extracting the global feature information during each second fusion processing.

5. The method of claim 4 , wherein obtaining the second global feature map with the lowest resolution by performing the at least one third fusion processing on associated feature points based on the first global feature map with the lowest resolution, comprises:
obtaining a third feature map by performing down-sampling processing on the first global feature map with the lowest resolution and by splicing image features of reduced feature points into image features of adjacent feature points;
obtaining a fourth feature map by performing at least one third fusion processing on associated feature points based on the third feature map;
obtaining a fifth feature map by performing up-sampling processing on the fourth feature map and by assigning image features of feature points in the fourth feature map to new adjacent feature points; and
obtaining the second global feature map with the lowest resolution by performing second fusion processing on associated feature points based on the fifth feature map and the first global feature map with the lowest resolution, and by extracting global feature information during the second fusion processing.

6. The method of claim 4 , wherein obtaining the at least one second global feature map by performing the at least one second fusion processing on associated feature points based on the second global feature map with the lowest resolution, and by extracting the global feature information during each second fusion processing, comprises:
obtaining, for each second fusion processing, a second global feature map corresponding to previous second fusion processing;
obtaining a sixth feature map by performing up-sampling processing on the second global feature map obtained and by assigning image features of feature points in the second global feature map to new adjacent feature points; and
obtaining a second global feature map corresponding to the second fusion processing by performing the second fusion processing on associated feature points based on the sixth feature map and a first global feature map having the same resolution as that of the sixth feature map, and by extracting the global feature information during the second fusion processing.

7. The method of claim 1 , wherein performing at least one image feature fusion processing on associated feature points based on the image to be segmented, and extracting the global feature information during each image feature fusion processing, comprises:
determining, based on an object type of the target object, window width and window level information corresponding to the image to be segmented;
obtaining a target segmentation image by adjusting pixel values of pixel points in the image to be segmented based on the window width and window level information; and
performing the at least one image feature fusion processing on associated feature points based on the target segmentation

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Yang Jingyi
• Yu Tan
• Sun Mingming
• Li Ping

Classifications

• G06V2201/03

  Recognition of patterns in medical or anatomical images · CPC title

• G06V10/42

  Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation · CPC title

• G06V10/7715

  Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods · CPC title

• G06V10/806

  of extracted features · CPC title

• G06V10/82

  using neural networks · CPC title