Key point detection method and device, equipment and storage medium

Abstract

The invention provides a key point detection method and device, equipment and a storage medium, and the method comprises the steps: obtaining a to-be-detected target image, and extracting the image features, which accord with thermodynamic diagram distribution, of the target image; performing down-sampling on the image features along a first direction and a second direction to obtain a first dimension reduction feature of the first direction and a second dimension reduction feature of the second direction; performing coordinate regression on the first dimension reduction feature to generate a first coordinate value, and performing coordinate regression on the second dimension reduction feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of a key point in the target image. According to the technical scheme provided by the invention, the data volume needing to be processed can be reduced while the detection precision is ensured.

Description

Technical Field

[0001] The present disclosure relates to the field of image processing technology, and in particular to a key point detection method, device, equipment and storage medium. Background Art

[0002] Key point detection is widely used in the field of computer vision. For example, in face recognition, virtual images, autonomous driving, etc., it is often necessary to detect key points on the face or body.

[0003] Currently, key point detection can be performed through heat map regression or coordinate regression. Among them, the heat map regression method can locate the key points in the image with high precision, but this method usually needs to process a large amount of data and has high requirements on the computing power of the device. The coordinate regression method needs to process a small amount of data, but the accuracy of key point detection is low and the detection error rate is high.

[0004] In view of this, a lightweight and highly accurate key point detection method is currently needed. Summary of the invention

[0005] In view of this, one or more embodiments of the present disclosure provide a key point detection method, apparatus, device and storage medium, which can reduce the amount of data to be processed while ensuring detection accuracy.

[0006] On the one hand, the present disclosure provides a key point detection method, which includes: acquiring a target image to be detected, and extracting image features of the target image that conform to the heat map distribution; downsampling the image features along a first direction and a second direction, respectively, to obtain a first reduced dimensionality feature in the first direction and a second reduced dimensionality feature in the second direction; performing coordinate regression on the first reduced dimensionality feature to generate a first coordinate value, and performing coordinate regression on the second reduced dimensionality feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

[0007] On the other hand, the present disclosure also provides a key point detection device, which includes: a feature extraction unit, used to obtain a target image to be detected, and extract image features of the target image that conform to the heat map distribution; a downsampling unit, used to downsample the image features along a first direction and a second direction, respectively, to obtain a first reduced dimensionality feature in the first direction and a second reduced dimensionality feature in the second direction; a coordinate regression unit, used to perform coordinate regression on the first reduced dimensionality feature to generate a first coordinate value, and to perform coordinate regression on the second reduced dimensionality feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

[0008] On the other hand, the present disclosure further provides an electronic device, which includes a memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned key point detection method is implemented.

[0009] On the other hand, the present disclosure further provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and when the computer program is executed by a processor, the above-mentioned key point detection method is implemented.

[0010] The technical solution provided by one or more embodiments of the present disclosure can downsample the image features in different directions after extracting the image features of the target image that conform to the heat map distribution, and perform coordinate regression on the reduced-dimensional features obtained by downsampling. Compared with the conventional heat map regression method, the computational overhead caused by directly processing high-resolution image features can be avoided. At the same time, compared with the conventional coordinate regression method, the accuracy of subsequent coordinate regression can be guaranteed due to the use of image features that conform to the heat map distribution. It can be seen that the technical solution provided by one or more embodiments of the present disclosure can reduce the amount of data that needs to be processed while ensuring detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The features and advantages of the embodiments of the present disclosure may be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present disclosure in any way. In the accompanying drawings:

[0012] Figure 1 A schematic diagram of a network architecture for heat map regression in related technologies is shown;

[0013] Figure 2 A schematic diagram of a network architecture for coordinate regression in related technologies is shown;

[0014] Figure 3 A flowchart of a method for key point detection in one embodiment of the present disclosure is shown;

[0015] Figure 4 A schematic diagram of a network architecture for key point detection in one embodiment of the present disclosure is shown;

[0016] Figure 5 A schematic diagram of functional modules of a key point detection device in one embodiment of the present disclosure is shown;

[0017] Figure 6 A schematic structural diagram of an electronic device in one embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0019] In the related art, the heat map regression method can be Figure 1 The architecture shown processes the input image to output the coordinates of the key points in the image. After downsampling and upsampling, the input image can obtain high-resolution features that conform to the heat map distribution. After the high-resolution features are activated by the normalization layer, the corresponding high-resolution probability map can be obtained. Subsequently, the high-resolution probability map is integrated through soft-argmax to output the horizontal and vertical coordinate values ​​of the key points.

[0020] In this process, high-resolution features are two-dimensional features with large data volumes. In the subsequent processing of high-resolution features, the computational overhead is large, which causes the solution to have high requirements on the computing power of the device and is not suitable for lightweight deployment scenarios such as mobile terminals.

[0021] In the related art, the coordinate regression method can be Figure 2 The architecture shown processes the input image to output the horizontal and vertical coordinate values ​​of the key points in the image. Among them, the coordinate regression method can directly process the low-resolution features after downsampling, thereby reducing the amount of data required to be processed. In this process, the low-resolution features contain less information, which may lead to inaccurate key point detection results.

[0022] In view of this, an embodiment of the present disclosure provides a key point detection method, please refer to Figure 3 , the method may include the following steps.

[0023] S1: Acquire a target image to be detected, and extract image features of the target image that conform to the heat map distribution.

[0024] Please combine Figure 4 In this embodiment, based on the heat map regression, the high-resolution features can be downsampled in different directions, and corresponding coordinate values ​​can be generated for different directions, thereby retaining the high detection accuracy of the heat map regression while reducing the amount of data required to be processed.

[0025] Specifically, the target image may be an image to be subjected to key point detection, and the content contained in the target image may be different according to different application scenarios. For example, in the scenario of face recognition, the content displayed in the target image may be a face. For another example, in the scenario of action posture recognition, the content displayed in the target image may be a human body.

[0026] In this embodiment, for the input target image, a method similar to heat map regression can be used to extract the image features of the target image that conform to the heat map distribution. The image features that conform to the heat map distribution can be Figure 4 High-resolution features in .

[0027] In practical applications, the feature extraction network can include downsampling and upsampling processes. In this way, for the input target image, the feature extraction network can first downsample the target image, and upsample the downsampled image features to obtain image features that conform to the heat map distribution. Among them, the downsampling process can be implemented in a variety of ways. For example, it can be implemented through a pooling layer or a convolutional layer. The upsampling process can also be implemented in a variety of ways. For example, it can be implemented through interpolation, transposed convolution or anti-pooling. Of course, in practical applications, network models such as HRNet or HourGlass can also be used to implement the feature extraction network.

[0028] It should be noted that the resolution of the extracted image features that conform to the heat map distribution is usually lower than the resolution of the original target image. For example, if the resolution of the original target image is H*W, then the resolution of the extracted image features that conform to the heat map distribution can be (H / s)*(W / s), where s can be used as the sampling ratio of the feature extraction network, which can be 4 or 8, for example.

[0029] In practical applications, the sampling ratio of the heat map regression method in the related art is usually set to 4. In one embodiment of the present disclosure, the above sampling ratio can be slightly larger than the sampling ratio in the conventional heat map regression method. In this way, the resolution of the image features extracted by the feature extraction network can be higher than Figure 2 Low resolution in and below Figure 1 High resolution in .

[0030] In one embodiment, the above-mentioned feature extraction network can be supervised by using heat map label features during offline training. Specifically, when training the feature extraction network, an image training sample can be obtained, and the image training sample can contain key points to be detected. At the same time, for the key points contained in the image training sample, matching heat map label features can be marked. Among them, the heat map label feature can be represented by a Gaussian heat map. In the heat map label feature, the closer the feature is to the key point, the higher the corresponding probability value. In this way, the heat map label feature as a whole can be expressed as a probability map that conforms to the Gaussian distribution.

[0031] The image training samples are downsampled by the feature extraction network to be trained, and the predicted image features are upsampled for the downsampled image features, and the predicted image features can be obtained. The predicted image features can be processed by a normalization layer to be converted into probability distribution features, wherein the normalization layer can be implemented by a softmax function. Subsequently, the probability distribution features and the heat map label features can be compared, and error information can be generated based on the comparison results. In practical applications, the error information between the probability distribution features and the heat map label features can be calculated based on MSE (Mean Squared Error) Loss. The error information can be corrected for the feature extraction network to be trained, so that under the supervision of the heat map label features, the corrected feature extraction network can output predicted image features that conform to the heat map distribution.

[0032] It should be noted that during the offline training phase, a normalization layer is connected after the feature extraction network to convert the predicted image features into probabilities. However, after the feature extraction network is trained, the normalization layer can be removed. Figure 4 The architecture shown in Figure 3 does not include this normalization layer.

[0033] S3: Downsample the image features along a first direction and a second direction respectively to obtain a first reduced-dimensionality feature in the first direction and a second reduced-dimensionality feature in the second direction.

[0034] In this embodiment, if the image features extracted in step S1 that meet the heat map distribution are directly processed, the amount of data to be processed is still large. In view of this, the extracted image features can be downsampled in different directions, and coordinate regression can be performed on the downsampled features, thereby reducing the amount of data to be processed during the coordinate regression process.

[0035] Specifically, the position of the key point needs to be determined in the target image, and the position of the key point can be represented by the coordinate value in the target image. Generally speaking, in a plane image, the coordinate value of the key point can be represented by a horizontal coordinate and a vertical coordinate. The horizontal coordinate and the vertical coordinate can both be the coordinates of the pixel point in the plane image.

[0036] In this embodiment, the image features can be downsampled along the first direction and the second direction, respectively, so as to obtain the first dimension reduction feature of the first direction and the second dimension reduction feature of the second direction. The first direction may be the X-axis direction, and the second direction may be the Y-axis direction. Of course, with the different application scenarios and the continuous advancement of technology, the target image is not necessarily a planar image. For example, the target image may be a captured or synthesized panoramic image. For a panoramic image, the first direction may be the longitude direction, and the second direction may be the latitude direction.

[0037] Taking the first direction as an example, when the extracted image features are downsampled along the first direction, several dimension reduction features in the first direction can be obtained, and the set of these dimension reduction features can together constitute the first dimension reduction feature in the first direction. Downsampling the image features in the first direction is equivalent to retaining the features in the second direction while ignoring or reducing the features in the first direction.

[0038] For example, the first direction is the X-axis direction. After the extracted image features are downsampled along the X-axis direction, several reduced-dimensional features distributed along the X-axis can be obtained. For each reduced-dimensional feature, the features in the X-axis direction are the same during the downsampling process. The downsampling process is actually to compress the features in the Y-axis direction of the image features and use the compressed features as reduced-dimensional features. In this way, the features in the Y-axis direction can be retained in the reduced-dimensional features.

[0039] In practical applications, assuming that the resolution of the image features extracted by the feature extraction network is (H / s)*(W / s), then the resolution of the first reduced dimension feature after downsampling along the X-axis direction can be (H / s)*1, and the resolution of the second reduced dimension feature after downsampling along the Y-axis direction can be 1*(W / s). It can be seen that compared with the image features extracted by the feature extraction network, the first reduced dimension feature and the second reduced dimension feature after downsampling have both shown a significant reduction in data volume.

[0040] It should be noted that in the above scenario, the first dimensionality reduction feature and the second dimensionality reduction feature after downsampling only sample one set of dimensionality reduction features in the corresponding downsampling direction. For example, the resolution (H / s)*1 indicates that only one set of dimensionality reduction features is sampled in the X-axis direction, and the set of dimensionality reduction features contains H / s features. However, in some application scenarios, the first dimensionality reduction feature and the second dimensionality reduction feature after downsampling do not necessarily sample only one set of dimensionality reduction features in the corresponding downsampling direction. For example, the resolution of the first dimensionality reduction feature after downsampling can be (H / s)*2, which means that two sets of dimensionality reduction features are sampled in the X-axis direction, and each set of dimensionality reduction features contains H / s features.

[0041] As for how many groups of dimensionality reduction features should be sampled in the first direction and the second direction respectively, it can be determined according to the requirements of the actual application scenario.

[0042] S5: Perform coordinate regression on the first dimensionality reduction feature to generate a first coordinate value, and perform coordinate regression on the second dimensionality reduction feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

[0043] In this embodiment, after downsampling the image features in different directions through different branches to obtain the first dimensionality reduction feature and the second dimensionality reduction feature, the first dimensionality reduction feature and the second dimensionality reduction feature can be respectively input into the fully connected layer, so as to perform coordinate regression through the fully connected layer to obtain the first coordinate value and the second coordinate value, respectively.

[0044] For example, after performing coordinate regression on the first dimension reduction feature obtained by downsampling along the X-axis direction, the coordinate value of the key point on the Y-axis can be obtained, and the coordinate value can be the first coordinate value mentioned above; similarly, after performing coordinate regression on the second dimension reduction feature obtained by downsampling along the Y-axis direction, the coordinate value of the key point on the X-axis can be obtained, and the coordinate value can be the second coordinate value mentioned above. In this way, the combination of (second coordinate value, first coordinate value) can represent the position of the key point in the target image.

[0045] In this embodiment, the downsampling and coordinate regression processes in steps S3 and S5 can be implemented by a coordinate regression network. The processing processes for different directions can be completed by different coordinate regression networks. The first coordinate regression network can downsample the image features along the first direction, and perform coordinate regression on the obtained first reduced dimension features, thereby outputting the first coordinate value. Similarly, the second coordinate regression network can downsample the image features along the second direction, and perform coordinate regression on the obtained second reduced dimension features, thereby outputting the second coordinate value.

[0046] When the above-mentioned first coordinate regression network or second coordinate regression network is trained offline, the heat map distribution feature can be obtained. The heat map distribution feature can be obtained by extracting features from image samples through a feature extraction network. At the same time, for the heat map distribution feature, key point coordinate information can be pre-marked, and the key point coordinate information can include a first coordinate score and a second coordinate score. For example, the first coordinate score can be an X coordinate, and the second coordinate score can be a Y coordinate.

[0047] Subsequently, the heat map distribution features can be downsampled along the first direction through the first coordinate regression network, and the obtained reduced dimension features can be subjected to coordinate regression to generate a first predicted coordinate value. Similarly, the heat map distribution features can be downsampled along the second direction through the second coordinate regression network, and the obtained reduced dimension features can be subjected to coordinate regression to generate a second predicted coordinate value.

[0048] Under the supervision of the first coordinate score and the second coordinate score, first error information can be generated based on the first predicted coordinate value and the first coordinate score, and the first coordinate regression network can be corrected using the first error information. And, second error information can be generated based on the second predicted coordinate value and the second coordinate score, and the second coordinate regression network can be corrected using the second error information. In practical applications, the first error information and the second error information can be generated by a cross entropy function.

[0049] In one embodiment, the target image may contain more than one key point, so that the number of the first coordinate value and the second coordinate value generated by the above steps is at least two. For example, the generated first coordinate value can be expressed as (y1, y2, y3, y4), and the second coordinate value can be expressed as (x1, x2, x3, x4). For the generated first coordinate value and the second coordinate value, the first coordinate value and the second coordinate value can be matched based on their respective generation order to divide the first coordinate value and the second coordinate value into multiple groups of coordinate pairs. For example, the first generated first coordinate value and the first generated second coordinate value can constitute a group of coordinate pairs. In this way, four groups of coordinate pairs (x1, y1), (x2, y2), (x3, y3), and (x4, y4) can be obtained. Among them, each group of coordinate pairs corresponds to each key point in the target image.

[0050] The technical solution provided by one or more embodiments of the present disclosure can downsample the image features in different directions after extracting the image features of the target image that conform to the heat map distribution, and perform coordinate regression on the reduced-dimensional features obtained by downsampling. Compared with the conventional heat map regression method, the computational overhead caused by directly processing high-resolution image features can be avoided. At the same time, compared with the conventional coordinate regression method, the accuracy of subsequent coordinate regression can be guaranteed due to the use of image features that conform to the heat map distribution. It can be seen that the technical solution provided by one or more embodiments of the present disclosure can reduce the amount of data that needs to be processed while ensuring detection accuracy.

[0051] See also Figure 5 In one embodiment of the present disclosure, a key point detection device is provided, the device comprising:

[0052] The feature extraction unit 100 is used to obtain a target image to be detected and extract image features of the target image that conform to the heat map distribution;

[0053] A downsampling unit 200 is used to downsample the image feature along a first direction and a second direction respectively to obtain a first reduced dimension feature in the first direction and a second reduced dimension feature in the second direction;

[0054] The coordinate regression unit 300 is used to perform coordinate regression on the first dimensionality reduction feature to generate a first coordinate value, and to perform coordinate regression on the second dimensionality reduction feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

[0055] Among them, the specific processing logic of each functional module can be found in the description of the aforementioned method implementation method, which will not be repeated here.

[0056] Each unit described in the above embodiments may be implemented by a computer chip or a product having a certain function. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0057] For the convenience of description, the above devices are described in terms of functions and are described separately in various units. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0058] See also Figure 6An embodiment of the present disclosure further provides an electronic device, which includes a memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned key point detection method is implemented.

[0059] The present disclosure also provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and when the computer program is executed by a processor, the above-mentioned key point detection method is implemented.

[0060] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0061] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implementing the method in the above method embodiment.

[0062] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0063] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above-mentioned types of memory.

[0064] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the embodiments of the device, equipment, and storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0065] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

[0066] Although the embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A key point detection method, characterized in that: The method comprises: Acquire a target image to be detected, and extract image features of the target image that conform to the thermal map distribution; Downsampling the image features along a first direction and a second direction respectively to obtain a first reduced-dimensionality feature in the first direction and a second reduced-dimensionality feature in the second direction; Performing coordinate regression on the first dimension reduction feature to generate a first coordinate value, and performing coordinate regression on the second dimension reduction feature to generate a second coordinate value; The combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

2. The method according to claim 1, characterized in that Extracting the image features of the target image that conform to the heat map distribution includes: The input target image is downsampled through a feature extraction network, and the downsampled image features are upsampled to obtain image features that conform to the heat map distribution.

3. The method according to claim 2, characterized in that The feature extraction network is trained based on the following method: Obtaining an image training sample, wherein the image training sample has a heat map label feature; Downsampling the image training samples through the feature extraction network to be trained, and upsampling the downsampled image features to obtain predicted image features; The predicted image features are converted into probability distribution features, and the probability distribution features are compared with the heat map label features, and error information is generated based on the comparison result, so as to correct the feature extraction network to be trained through the error information.

4. The method according to claim 3, characterized in that Converting the predicted image features into probability distribution features includes: The predicted image features are input into a normalization layer, so as to output probability distribution features corresponding to the predicted image features through the normalization layer.

5. The method according to claim 1, characterized in that The image features are downsampled along a first direction through a first coordinate regression network, and coordinate regression is performed on the obtained first reduced-dimensional features to generate first coordinate values; the image features are downsampled along a second direction through a second coordinate regression network, and coordinate regression is performed on the obtained second reduced-dimensional features to generate second coordinate values.

6. The method according to claim 5, characterized in that The first coordinate regression network and the second coordinate regression network are trained based on the following method: Acquire a heat map distribution feature, wherein the heat map distribution feature has marked key point coordinate information, and the key point coordinate information includes a first coordinate score and a second coordinate score; Downsampling the heat map distribution features along the first direction through the first coordinate regression network, and performing coordinate regression on the obtained reduced-dimensional features to generate a first predicted coordinate value; Downsampling the heat map distribution features along the second direction through the second coordinate regression network, and performing coordinate regression on the obtained reduced-dimensional features to generate second predicted coordinate values; generating first error information based on the first predicted coordinate value and the first coordinate score, and correcting the first coordinate regression network using the first error information; Based on the second predicted coordinate value and the second coordinate score, second error information is generated, and the second coordinate regression network is corrected using the second error information.

7. The method according to claim 1, characterized in that The number of the first coordinate value and the number of the second coordinate value are both at least two; after generating the first coordinate value and the second coordinate value, the method further includes: Based on the generation order of the first coordinate value and the second coordinate value, the first coordinate value and the second coordinate value are matched to divide the first coordinate value and the second coordinate value into multiple groups of coordinate pairs; wherein each group of coordinate pairs corresponds to each key point in the target image.

8. A key point detection device, characterized in that: The device comprises: A feature extraction unit, used to obtain a target image to be detected and extract image features of the target image that conform to the thermal map distribution; A downsampling unit, configured to downsample the image feature along a first direction and a second direction respectively, to obtain a first reduced dimension feature in the first direction and a second reduced dimension feature in the second direction; A coordinate regression unit is used to perform coordinate regression on the first dimensionality reduction feature to generate a first coordinate value, and to perform coordinate regression on the second dimensionality reduction feature to generate a second coordinate value; wherein the combination of the first coordinate value and the second coordinate value represents the position of the key point in the target image.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images