Method and apparatus for detecting target point in image, and computer storage medium
View Patent ↗Embodiments of this application provide a method for detecting a target point in an image. The method may include: obtaining an image under test, where the image under test may include a structure-stable first target object and a structure-unstable second target object; and processing the image under test based on a target point detection model to obtain a target point in the image under test, where the target point may include a feature point on the structure-stable first target object and a feature point on the structure-unstable second target object.
1. A method for detecting a target point in an image, characterized in that the method comprises:
obtaining an image under test, wherein the image under test comprises a structure-stable first target object and a structure-unstable second target object;
processing the image under test based on a target point detection model to obtain a target point in the image under test, wherein the target point comprises a feature point on the structure-stable first target object and a feature point on the structure-unstable second target object; and
training a parameter of the target point detection model based on a sample image and labeled data, wherein the sample image comprises the first target object and the second target object, and the labeled data comprises a labeled feature point on the first target object and a labeled feature point on the second target object,
wherein the training a parameter of the target point detection model based on a sample image and labeled data comprises:
inputting the sample image into a to-be-trained target point detection model to obtain a predicted target point of the sample image;
determining a loss value based on the predicted target point and the labeled data; and
adjusting the parameter of the target point detection model until the loss value satisfies a first preset loss value to obtain a parameter of the trained target point detection model.
2. The method according to claim 1 , characterized in that the first target object has a structure-stable specified structure, and a partial region of the second target object has a structure-stable specified structure.
3. The method according to claim 1 , characterized in that the method further comprises:
the labeled feature point on the first target object is at least one coordinate point on the first target object; and
the labeled feature point on the second target object is at least one coordinate point on a region of the second target object having a specified structure.
4. The method according to claim 1 , characterized in that before the training a parameter of the target point detection model based on a sample image and labeled data, the method further comprises:
obtaining the labeled data based on the sample image and an initial target point detection model.
5. The method according to claim 4 , characterized in that the obtaining the labeled data based on the sample image and an initial target point detection model comprises:
inputting the sample image into the initial target point detection model to obtain an initial predicted target point of the sample image;
determining an initial loss value based on the initial predicted target point and initial labeled data; and
adjusting the initial labeled data to make the initial loss value satisfy a second preset loss value, so as to obtain the labeled data.
6. The method according to claim 1 , characterized in that the image under test is an image of an adapting piece and a pole that are welded in a battery cell, the first target object is the pole, and the second target object is the adapting piece.
7. The method according to claim 6 , characterized in that the structure-stable specified structure of the first target object comprises a groove at a center of the pole.
8. The method according to claim 6 , characterized in that the region of the second target object having the structure-stable specified structure comprises a tail end of the adapting piece and a region close to a second bend adjacent to the tail end of the adapting piece.
9. The method according to claim 8 , characterized in that the labeled feature point on the first target object is a coordinate point of an end point of the groove at the center of the pole; and the labeled feature point on the second target object comprises a coordinate point of the tail end of the adapting piece and a coordinate point in the region close to the second bend adjacent to the tail end of the adapting piece.
10. The method according to claim 1 , characterized in that the target point detection model is a residual neural network.
11. An apparatus for detecting a target point in an image, characterized by comprising a processor and a memory, wherein the memory is configured to store a program, and the processor is configured to call and run the program to execute the method for detecting a target point in an image according to claim 1 .
12. A non-transitory computer-readable storage medium, characterized by comprising a computer program, wherein when the computer program runs on a computer, the computer executes the method for detecting a target point in an image according to claim 1 .
13. An apparatus for detecting a target point in an image, characterized in that the apparatus comprises:
an obtaining circuitry, configured to obtain an image under test, wherein the image under test comprises a structure-stable first target object and a structure-unstable second target object;
a processing circuitry, configured to process the image under test based on a target point detection model to obtain a target point in the image under test, wherein the target point comprises a feature point on the structure-stable first target object and a feature point on the structure-unstable second target object; and
a training circuitry, configured to train a parameter of the target point detection model based on a sample image and labeled data, wherein the sample image comprises the first target object and the second target object, and the labeled data comprises a labeled feature point on the first target object and a labeled feature point on the second target object,
wherein the training circuitry further comprises:
a first inputting circuitry, configured to input the sample image into a to-be-trained target point detection model to obtain a predicted target point of the sample image;
a first determining circuitry, configured to determine a loss value based on the predicted target point and the labeled data; and
a first adjusting circuitry, configured to adjust the parameter of the target point detection model until the loss value satisfies a first preset loss value to obtain a parameter of the trained target point detection model.
14. The apparatus according to claim 13 , characterized in that the first target object has a specified structure, and a partial region of the second target object has a structure-stable specified structure.
15. The apparatus according to claim 13 , characterized in that the labeled feature point on the first target object is at least one coordinate point on the first target object; and the labeled feature point on the second target object is at least one coordinate point on a region of the second target object having a specified structure.
16. The apparatus according to claim 13 , characterized in that the apparatus further comprises:
a data labeling circuitry, configured to: before the training a parameter of the target point detection model based on a sample image and labeled data, obtain the labeled data based on the sample image and an initial target point detection model.
17. The apparatus according to claim 16 , characterized in that the data labeling module further comprises:
a second inputting circuitry, configured to input the sample image into the initial target point detection model to obtain an initial predicted target point of the sample image;
a second determining circuitry, configured to determine an initial loss value based on the initial predicted target point and initial labeled data; and
a second adjusting circuitry, configured to adjust the initial labeled data to make the initial loss value satisfy a second preset loss value, so as to obtain the labeled data.
18. The apparatus according to claim 13 , characterized in that the image under test is an image of an adapting piece and a pole that are welded in a battery cell, the first target object is the pole, and the second target object is the adapting piece.
19. The apparatus according to claim 18 , characterized in that the structure-stable specified structure of the first target object comprises a groove at a center of the pole.
20. The apparatus according to claim 18 , characterized in that the region of the second target object having the structure-stable specified structure comprises a tail end of the adapting piece and a region close to a second bend adjacent to the tail end of the adapting piece.
21. The apparatus according to claim 20 , characterized in that the labeled feature point on the first target object is a coordinate point of an end point of the groove at the center of the pole; and the labeled feature point on the second target object comprises a coordinate point of the tail end of the adapting piece and a coordinate point in the region close to the second bend adjacent to the tail end of the adapting piece.
22. The apparatus according to claim 13 , characterized in that the target point detection model is a residual neural network.