Granted Patent B2

US 12,567,224 · App. 18/361,972 · Granted Mar 3, 2026

Inventors: Boxiong Huang (Ningde, CN); Guannan Jiang (Ningde, CN); Zhiyu Wang (Ningde, CN)

Assignee: CONTEMPORARY AMPEREX TECHNOLOGY (HONG KONG) LIMITED

G06V10/25G06T7/50G06T7/73G06V10/44G06V20/70G06T2207/20084G06V2201/07

Abstract

Claims (56)

1 . A method for locating a target region on an image, comprising:

obtaining coordinates of target points in a to-be-detected image based on a feature point recognition model, wherein the to-be-detected image comprises an unstably imaged target region, a target object in the target region comprises a stably imaged first section and an unstably imaged second section, and the target points comprise a feature point on the first section; and

locating the target region based on the coordinates of the target points;

wherein before obtaining coordinates of the target points in the to-be-detected image based on the feature point recognition model, the method further comprises:

training the feature point recognition model based on training data, wherein the training data comprises labeled feature points, and the labeled feature points comprise the feature point on the first section; and

wherein training the feature point recognition model based on the training data comprises:

obtaining the training data, wherein the training data comprises a training image set and a feature point dataset:

inputting the training data into a to-be-trained neural network model to perform feature extraction and obtain an estimate of coordinates of the feature point:

determining a function value of a first loss function based on the estimate of the coordinates of the feature point and a ground truth corresponding to the estimate of the coordinates of the feature point in the training dataset, wherein the function value of the first loss function is used to indicate a deviation between the estimate of the coordinates of the feature point and the ground truth;

updating a parameter of the to-be-trained neural network model based on the function value of the first loss function; and

determining that the function value of the first loss function has converged to a first threshold, and then using a current neural network model as the feature point recognition model.

2 . The method according to claim 1 , wherein obtaining the training data comprises:

generating a training image based on a depth map that comprises the target region; and

labeling feature points for the training image to obtain the training image set.

3 . The method according to claim 1 , wherein obtaining the training data comprises:

sorting feature points on a training image to obtain the feature point dataset.

4 . The method according to claim 1 , wherein the feature point dataset comprises the coordinates and a serial number of the feature point.

5 . The method according to claim 1 , wherein the feature point dataset comprises a descriptor of the feature point and a relative positional relationship between feature points.

6 . The method according to claim 1 , wherein before obtaining coordinates of target points in the to-be-detected image based on the feature point recognition model, the method further comprises:

obtaining an original image; and

converting the original image into an image in a PNG format to obtain the to-be-detected image.

7 . The method according to claim 1 , wherein the feature point dataset is stored in a JSON format.

8 . The method according to claim 1 , wherein, after determining that the function value of the first loss function has converged to the first threshold, the method further comprises:

inputting the training image set as a test set into the feature point recognition model to obtain an estimate of the feature point; and

calculating an accuracy metric based on the estimate of the feature point and the ground truth of the coordinates of the feature point corresponding to the training image set.

9 . The method according to claim 8 , wherein the accuracy metric is a Euclidean square root mean value of the estimate of the feature point and the ground truth of the coordinates of the feature point corresponding to the training image set.

10 . The method according to claim 1 , wherein the to-be-trained neural network model comprises a backbone feature extraction neural network, and the backbone feature extraction neural network is a ResNet 18 neural network.

11 . The method according to claim 1 , wherein the first loss function is a Smooth L1 loss function.

12 . The method according to claim 1 , wherein

the to-be-detected image is an image of an adapter strip and an electrode post that are welded in a battery cell; and

the target region comprises a weld region of the electrode post and the adapter strip.

13 . The method according to claim 12 , wherein the target object is the adapter strip, the adapter strip comprises a tab connecting portion and an electrode post connecting portion, one end of the tab connecting portion is connected to the electrode post connecting portion, the first section is the tab connecting portion, and the target points comprise coordinate points located at two ends of the tab connecting portion.

14 . A device, comprising a processor and a memory;

wherein the memory is configured to store a program, and the processor is configured to call the program from the memory and run the program;

wherein the program, when executed, causes the device to perform a method for locating a target region on an image that comprises:

obtaining coordinates of target points in a to-be-detected image based on a feature point recognition model, wherein the to-be-detected image comprises an unstably imaged target region, a target object in the target region comprises a stably imaged first section and an unstably imaged second section, and the target points comprise a feature point on the first section; and

locating the target region based on the coordinates of the target points;

wherein before obtaining coordinates of the target points in the to-be-detected image based on the feature point recognition model, the method further comprises:

training the feature point recognition model based on training data, wherein the training data comprises labeled feature points, and the labeled feature points comprise the feature point on the first section; and

wherein training the feature point recognition model based on the training data comprises:

obtaining the training data, wherein the training data comprises a training image set and a feature point dataset:

inputting the training data into a to-be-trained neural network model to perform feature extraction and obtain an estimate of coordinates of the feature point:

determining a function value of a first loss function based on the estimate of the coordinates of the feature point and a ground truth corresponding to the estimate of the coordinates of the feature point in the training dataset, wherein the function value of the first loss function is used to indicate a deviation between the estimate of the coordinates of the feature point and the ground truth;

updating a parameter of the to-be-trained neural network model based on the function value of the first loss function; and

determining that the function value of the first loss function has converged to a first threshold, and then using a current neural network model as the feature point recognition model.

15 . A non-transitory computer-readable storage medium, configured to store a computer program, which when executed on a computer, causes the computer to perform a method for locating a target region on an image that comprises:

obtaining coordinates of target points in a to-be-detected image based on a feature point recognition model, wherein the to-be detected image comprises an unstably imaged target region, a target object in the target region comprises a stably imaged first section and an unstably imaged second section, and the target points comprise a feature point on the first section; and

locating the target region based on the coordinates of the target points;

wherein before obtaining coordinates of the target points in the to-be-detected image based on the feature point recognition model, the method further comprises:

training the feature point recognition model based on training data, wherein the training data comprises labeled feature points, and the labeled feature points comprise the feature point on the first section; and

wherein training the feature point recognition model based on the training data comprises:

obtaining the training data, wherein the training data comprises a training image set and a feature point dataset:

inputting the training data into a to-be-trained neural network model to perform feature extraction and obtain an estimate of coordinates of the feature point:

determining a function value of a first loss function based on the estimate of the coordinates of the feature point and a ground truth corresponding to the estimate of the coordinates of the feature point in the training dataset, wherein the function value of the first loss function is used to indicate a deviation between the estimate of the coordinates of the feature point and the ground truth:

updating a parameter of the to-be-trained neural network model based on the function value of the first loss function; and

determining that the function value of the first loss function has converged to a first threshold, and then using a current neural network model as the feature point recognition model.

Assignments (2)

Priority Claims (1)

Continuity (2)

References Cited (36)