Method and apparatus for evaluating image relative definition, device and medium

Patent metadata

Field	Value
Publication number	US-11921276-B2
Application number	US-202117379428-A
Country	US
Kind code	B2
Filing date	Jul 19, 2021
Priority date	Oct 23, 2020
Publication date	Mar 5, 2024
Grant date	Mar 5, 2024

Abstract

Provided are a method and apparatus for evaluating image relative definition, a device and a medium, relating to technologies such as computer vision, deep learning and intelligent medical. A specific implementation solution is: extracting a multi-scale feature of each image in an image set, where the multi-scale feature is used for representing definition features of objects having different sizes in an image; and scoring relative definition of each image in the image set according to the multi-scale feature by using a relative definition scoring model pre-trained, where the purpose for training the relative definition scoring model is to learn a feature related to image definition in the multi-scale feature.

First claim

What is claimed is:

1. A method for evaluating image relative definition, comprising:
extracting a multi-scale feature of each image in an image set, wherein the multi-scale feature is used for representing definition features of objects having different sizes in the each image; and
scoring relative definition of the each image in the image set according to the multi-scale feature by using a pre-trained relative definition scoring model,
wherein the pre-trained relative definition scoring model is obtained by training a pre-established relative definition scoring model, and the pre-trained relative definition scoring model is used to learn a feature related to image definition in a multi-scale feature;
wherein extracting the multi-scale feature of each image in the image set comprises:
extracting the multi-scale feature of each image in the image set by using a pyramid Laplace operator, wherein the pyramid Laplace operator is used for respectively extracting features by using the pyramid Laplace operator at a plurality of image scales and merging the extracted features; and
wherein extracting the multi-scale feature of each image in the image set by using the pyramid Laplace operator comprises:
scaling each image in the image set according to at least two multiple values to obtain at least two scaled images corresponding to the each image;
extracting a multi-scale feature of each of the at least two scaled images by using the pyramid Laplace operator; and
merging the multi-scale features of the at least two scaled images corresponding to the each image to obtain the multi-scale feature of the each image.

2. The method of claim 1 , wherein the pre-established relative definition scoring model is trained in a contrastive training manner.

3. The method of claim 1 , wherein the images in the image set are captured by a capturing device at different focal lengths in a same field of view.

4. The method of claim 3 , wherein a process for training the pre-established relative definition scoring model comprises:
sequencing sample images in a sample image set according to magnitudes of capturing focal lengths;
determining, in the sample image set after the sequencing, at least one clear sample image labeled;
determining a plurality of paired sample image groups and a relative definition relationship between two sample images in each of the plurality of sample image groups according to the at least one clear sample image and the sequencing of the sample images in the sample image set;
inputting each of the plurality of sample image groups as training data into the pre-established relative definition scoring model, wherein the pre-established relative definition scoring model is configured to determine a relative definition score of each of the two sample images according to multi-scale features of the two sample images in the each of the plurality of sample image groups; and
training the pre-established relative definition scoring model according to the relative definition relationship and the relative definition scores of the two sample images in the each of the plurality of sample image groups by using a contrastive loss function.

5. The method of claim 4 , wherein the pre-established relative definition scoring model is a twin network structure, network branches are configured to score relative definition of a sample image having higher definition and a sample image having lower definition which are in the each of the plurality of sample image groups.

6. The method of claim 1 , further comprising:
performing definition sequencing on the images in the image set based on scoring results.

7. The method of claim 1 , wherein the images are microscopic images.

8. An apparatus for evaluating image relative definition, comprising:
a multi-scale feature extraction module, which is configured to extract a multi-scale feature of each image in an image set, wherein the multi-scale feature is used for representing definition features of objects having different sizes in the each image; and
a relative definition scoring module, which is configured to score relative definition of the each image in the image set according to the multi-scale feature by using a pre-trained relative definition scoring model,
wherein the pre-trained relative definition scoring model is obtained by training a pre-established relative definition scoring model, and the pre-trained
relative definition scoring model is used to learn a feature related to image definition in the multi-scale feature;
wherein the multi-scale feature extraction module is configured to:
extract the multi-scale feature of each image in the image set by using a pyramid Laplace operator, wherein the pyramid Laplace operator is used for extracting features by respectively using the pyramid Laplace operator at a plurality of image scales and merging the extracted features; and
wherein the multi-scale feature extraction module comprises:
an image scaling unit, which is configured to scale each image in the image set according to at least two multiple values to obtain at least two scaled images corresponding to the each image;
a feature extraction unit, which is configured to extract a multi-scale feature of each of the at least two scaled images by using the pyramid Laplace operator; and
a feature merging unit, which is configured to merge the multi-scale features of the at least two scaled images corresponding to the each image to obtain the multi-scale feature of the each image.

9. The apparatus of claim 8 , wherein the pre-established relative definition scoring model is trained in a contrastive training manner.

10. The apparatus of claim 8 , wherein images in the image set are captured by a capturing device at different focal lengths in a same field of view.

11. The apparatus of claim 10 , further comprising: a relative definition scoring model training module, which comprises:
a sequencing unit, which is configured to sequence sample images in a sample image set according to magnitudes of capturing focal lengths;
a labeling determination unit, which is configured to determine, in the sample image set after the sequencing, at least one clear sample image labeled;
an image group determination unit, which is configured to determine a plurality of paired sample image groups and a relative definition relationship between two sample images in each of the plurality of sample image groups according to the at least one clear sample image and the sequencing of the sample images in the sample image set;
an input unit, which is configured to input each of the plurality of sample image groups as training data into pre-established relative definition scoring model, wherein the pre-established relative definition scoring model is configured to determine a relative definition score of each of the two sample image according to multi-scale features of the two sample images in the each of the plurality of sample image groups; and
a training unit, which is configured to train the pre-established relative definition scoring model according to the relative definition relationship and the relative definition scores of the two sample images in the each of the plurality of sample image groups by using a contrastive loss function.

12. The apparatus of claim 11 , wherein the pre-established relative definition scoring model is a twin network structure, network branches are respectively configured to score relative definition of a sample image having higher definition and a sample image having lower definition which are in the each of the plurality of sample image groups.

13. The apparatus of claim 8 , further comprising:
a

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Long Xiang
• Peng Yan
• Lin Shufei
• Xin Ying
• Zhang Bin
• Yuan Pengcheng
• Wang Xiaodi
• Feng Yuan
• Han Shumin

Classifications

• G02B21/244 Primary

  using image analysis techniques · CPC title

• G02B21/367

  providing an output produced by processing a plurality of individual source images, e.g. image tiling, montage, composite images, depth sectioning, image comparison · CPC title

• G06F18/213

  Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06T7/0002 Primary

  Inspection of images, e.g. flaw detection · CPC title