Method, apparatus and system for simulated generation of palm feature data, and storage medium

By constructing a palm feature learning database and training a denoising network, and combining it with intra-class augmentation techniques to generate high-quality and highly diverse palm feature data, the problem of scarce palm feature data is solved, and the robustness of the palm recognition model is improved.

WO2026232842A1PCT designated stage Publication Date: 2026-11-12GRG BANKING EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/146906
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-07
Filing Date
2025-12-30
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

The high cost and scarcity of palm feature data acquisition in existing technologies limit the research and development of palm feature recognition algorithms.

Method used

A palm feature learning database is constructed. Initial palm feature images are generated by training a denoising network, and target palm feature images are generated by in-class augmentation to simulate palm feature changes under various real-world conditions.

Benefits of technology

It enables the batch generation of high-quality and highly diverse palm feature data, improves the robustness of the palm recognition model, and meets the training requirements of the palm feature recognition model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025146906_12112026_PF_FP_ABST
    Figure CN2025146906_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A method, apparatus and system for simulated generation of palm feature data, and a storage medium. The method for simulated generation of palm feature data comprises: constructing a palm feature learning database; on the basis of the palm feature learning database, training a plurality of denoising networks; on the basis of the plurality of trained denoising networks, generating a plurality of initial palm feature images in batches; and performing intra-class augmentation on the plurality of initial palm feature images, so as to obtain a plurality of target palm feature images.
Need to check novelty before this filing date? Find Prior Art

Description

Method, device, system and storage medium for simulating generation of palm feature data

[0001] Priority information

[0002] This application claims priority to and the benefit of the filing date of Chinese Patent Application No. 202510581438.1 filed on May 7, 2025 in the China National Intellectual Property Office, and specifically claims priority to and the benefit of the filing date of the aforementioned application, and incorporates herein by reference in its entirety for all purposes. TECHNICAL FIELD

[0003] The present application relates to the technical field of image processing, and particularly relates to a method, device, system and computer readable storage medium for simulating generation of palm feature data. BACKGROUND

[0004] Palm feature recognition is a new biometric recognition technology, which has better privacy and security than conventional face recognition technology. However, there is very little publicly available palm feature data at present, and the cost of collecting palm feature data is very high, which limits the research and development of related recognition algorithms. Under this condition, how to generate a large amount of new palm feature data is a problem to be solved at present. SUMMARY

[0005] The present application provides a method, device, system and computer readable storage medium for simulating generation of palm feature data to solve at least one of the above technical problems.

[0006] The method for simulating generation of palm feature data according to the present application embodiment comprises:

[0007] constructing a palm feature learning database;

[0008] training a plurality of denoising networks based on the palm feature learning database;

[0009] generating a plurality of initial palm feature images in batches based on the plurality of trained denoising networks;

[0010] performing intra-class augmentation on the plurality of initial palm feature images to obtain a plurality of target palm feature images.

[0011] The device for simulating generation of palm feature data according to the present application embodiment comprises:

[0012] a construction module configured to construct a palm feature learning database;

[0013] a training module configured to train a plurality of denoising networks based on the palm feature learning database;

[0014] The generation module is used to generate multiple initial palm feature images in batches based on the multiple trained denoising networks;

[0015] An expansion module is used to expand multiple initial palm feature images within a class to obtain multiple target palm feature images.

[0016] The palm feature data simulation generation system of this application includes one or more processors and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the above-described palm feature data simulation generation method.

[0017] The computer-readable storage medium of this application embodiment stores a computer program thereon, which, when executed by a processor, implements the above-described method for simulating and generating palm feature data.

[0018] Additional aspects and advantages of embodiments of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:

[0020] Figure 1 is a flowchart illustrating a method for simulating and generating palm feature data according to certain embodiments of this application;

[0021] Figure 2 is a schematic diagram of an application scenario of the palm feature data simulation generation method according to certain embodiments of this application;

[0022] Figure 3 is a flowchart illustrating a method for simulating and generating palm feature data according to certain embodiments of this application;

[0023] Figure 4 is a flowchart illustrating the training of a denoising network according to certain embodiments of this application;

[0024] Figure 5 is a flowchart illustrating a method for simulating and generating palm feature data according to certain embodiments of this application;

[0025] Figure 6 is a flowchart illustrating a method for simulating and generating palm feature data according to certain embodiments of this application;

[0026] Figure 7 is a flowchart illustrating a method for simulating and generating palm feature data according to certain embodiments of this application;

[0027] Figure 8 is a flowchart illustrating the intra-class expansion of an initial palm feature image according to certain embodiments of this application.

[0028] Figure 9 is a schematic diagram of a module for simulating and generating palm feature data according to certain embodiments of this application;

[0029] Figure 10 is a schematic diagram of a device for simulating and generating palm feature data according to certain embodiments of this application;

[0030] Figure 11 is a schematic diagram of the modules of a palm feature data simulation generation system according to certain embodiments of this application;

[0031] Figure 12 is a schematic diagram of the connection state between a computer-readable storage medium and a processor according to certain embodiments of this application. Embodiments of the present invention

[0032] The embodiments of this application will be further described below with reference to the accompanying drawings. The same or similar reference numerals in the drawings denote the same or similar elements or elements having the same or similar functions throughout. Furthermore, the embodiments of this application described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of this application, and should not be construed as limiting this application.

[0033] Please refer to Figure 1. This application provides a method for simulating and generating palm feature data. The method for simulating and generating palm feature data includes:

[0034] 010: Construct a palm feature learning database;

[0035] 020: Training multiple denoising networks based on a palm feature learning database;

[0036] 030: Generate multiple initial palm feature images in batches based on multiple trained denoising networks;

[0037] 040: Perform intra-class expansion on multiple initial palm feature images to obtain multiple target palm feature images.

[0038] In the method for simulating and generating palm feature data according to the embodiments of this application, multiple denoising networks are trained based on a palm feature learning database, and multiple initial palm feature images are generated in batches based on the trained multiple denoising networks. Then, intra-class expansion is performed on the multiple initial palm feature images to obtain multiple target palm feature images. In this way, high-quality and highly diverse palm feature data can be generated in batches, which can meet the training requirements of the palm feature recognition model and is beneficial to improving the robustness of the palm recognition model.

[0039] In some implementations, a palm feature learning database is constructed, which includes real palm feature data. Multiple denoising networks can be trained based on this database. These denoising networks utilize deep learning techniques to reduce or eliminate noise in images. By training multiple denoising networks, they can acquire the ability to generate high-quality palm feature data from random noise. The denoising networks can be constructed based on a U-Net architecture.

[0040] After multiple denoising networks are trained, random noise can be processed by multiple denoising networks in sequence to obtain an initial palm feature image. Thus, multiple initial palm feature images can be generated in batches from one or more random noises based on multiple denoising networks.

[0041] After obtaining multiple initial palm feature images, intra-class augmentation can be performed on these initial palm feature images to obtain multiple target palm feature images, thus ensuring the diversity of the simulated palm feature data. The intra-class augmentation strategy can be determined according to the actual application requirements.

[0042] Taking the generated target palm feature image as an example for training a palm feature recognition model, since the palm feature recognition model will be affected by factors such as palm movement and tilting in actual application, it is expected that when training the palm feature recognition model, the ability of the palm feature recognition model to recognize moving and tilted palm features will be trained.

[0043] Therefore, during intra-class expansion, a corresponding expansion strategy can be determined to simulate real hand movement and tilting. This allows the target hand feature image obtained after intra-class expansion to simulate the features of a moving and tilted hand. Training the hand feature recognition model based on this target hand feature image can improve the model's ability to recognize moving and tilted hand features. This satisfies the training requirements of the hand feature recognition model and improves its robustness.

[0044] It should be noted that palm features can include palm vein features and palm print features, and the simulation generation method of this application can support the simulation generation of palm vein feature data and palm print feature data. The palm feature learning database can include both palm vein feature data and palm print feature data. However, in the training of multiple denoising networks based on the palm feature learning database and in subsequent processes, it is necessary to further distinguish between palm vein feature data and palm print feature data.

[0045] For example, multiple denoising networks can be trained based on palm vein feature data in a palm feature learning database; multiple initial palm feature images including palm vein features can be generated in batches based on the trained multiple denoising networks; and intra-class expansion can be performed on the multiple initial palm feature images including palm vein features to obtain multiple target palm feature images including palm vein features.

[0046] For example, multiple denoising networks can be trained based on palm print feature data in a palm feature learning database; multiple initial palm feature images including palm print features can be generated in batches based on the trained multiple denoising networks; and intra-class expansion can be performed on the multiple initial palm feature images including palm print features to obtain multiple target palm feature images including palm print features.

[0047] The following describes in detail the specific process of constructing a feature learning database.

[0048] Referring to Figure 2, in some embodiments, the palm feature learning database includes feature learning images. Constructing the palm feature learning database (i.e., 010) includes:

[0049] 011: Obtain a realistic, complete image of the palm;

[0050] 012: Perform key point detection on the complete palm image;

[0051] 013: Based on the detected key points, perform image angle correction on the complete palm image to obtain the corrected complete palm image;

[0052] 014: Cropping the region of interest from the corrected complete hand image yields the feature-learned image.

[0053] In some implementations, images of a real hand are acquired to obtain multiple complete hand images. These multiple complete hand images can be divided into palm print images and palm vein images, facilitating subsequent training of the denoising network separately. Keypoint detection is performed on each complete hand image using a detection model. Keypoint detection refers to detecting key points on the palm within the complete hand image. Any keypoint detection model can be used, and no restrictions are imposed here.

[0054] By detecting key points, multiple key points can be obtained for each complete hand image. Based on these key points, the position and angle of the hand in the corresponding complete hand image can be determined. Since the hand may move or tilt during image acquisition, resulting in a tilt in the complete hand image, the corresponding complete hand image is corrected based on the detected key points to obtain a corrected complete hand image. This corrected image ensures that the tilt angle of the hand is zero, effectively straightening the hand.

[0055] This facilitates the subsequent training of the denoising network, making it easier for the denoising network to learn the essential features of the palm (such as texture and geometric structure) rather than irrelevant pose differences, resulting in more realistic and structurally reasonable images generated by the denoising network.

[0056] After image angle correction, the region of interest (ROI) is determined in each corrected complete hand image. The ROI in the corrected complete hand image is the region where the hand features are located. Cropping the ROI of each corrected complete hand image yields the corresponding feature learning image. This facilitates the subsequent training of the denoising network, avoiding interference from irrelevant information in the image and affecting the training results.

[0057] Multiple feature-learning images can form a hand feature learning database. After obtaining the hand feature learning database, multiple denoising networks can be trained based on it. The training process of the denoising network is described in detail below.

[0058] Referring to Figures 3 and 4, in some implementations, multiple denoising networks are trained (i.e., O2O) based on a palm feature learning database, including:

[0059] 021: Determine training images from the palm feature learning database;

[0060] 022: Based on the forward diffusion process of the diffusion model, random Gaussian noise is added to the training image one by one, so that the training image changes from the initial noise state to the result noise state;

[0061] 023: The reverse diffusion process based on the diffusion model successively denoises the training image of the result noise state through a denoising network in order to train multiple denoising networks.

[0062] In some implementations, after learning from the palm feature database, training images are determined from the palm feature learning database. If the network being trained is a denoising network that generates palm vein feature data, then multiple feature learning images including palm vein features are randomly selected from the palm feature learning database as training images. If the network being trained is a denoising network that generates palmprint feature data, then multiple feature learning images including palmprint features are randomly selected from the palm feature learning database as training images. There is no limit to the number of training images.

[0063] Using the noise state of the training image as the initial noise state, This is represented as follows. Then, based on the forward diffusion process of the diffusion model, random Gaussian noise is successively added to the training image. After adding random Gaussian noise to the training image once, the training image changes from the initial noise state to the first noise state, as expressed below:

[0064]

[0065] in, This indicates the first noise state. The parameter value to add noise at the set time zero. Random Gaussian noise was added at time zero. And so on, after... The noise addition operation transforms the training image from its initial noisy state to its final noisy state, as shown in the following expression:

[0066]

[0067] in, express The noise state of the result at any given time. express The corresponding time Noise status. For setting The parameter values ​​for adding noise at any given time. for The random Gaussian noise added at time t. The resulting noise state can also be expressed by the following formula:

[0068]

[0069] in, For the predefined noise parameter planned values, =1, 2, 3, ... Based on the planned values, the initial noise state can be obtained. To the result noise state Any intermediate result. for The total random Gaussian noise added each time.

[0070] Subsequently, based on the back-diffusion process of the diffusion model, the training image in the resulting noisy state is successively denoised through the denoising network, so that the training image is transformed from the resulting noisy state to the initial noisy state, in order to train multiple denoising networks. Each denoising network is used to predict the noise to be eliminated in one state transition. From the resulting noisy state to the initial noisy state, multiple state transitions are required. Therefore, multiple denoising networks need to be trained.

[0071] Referring to Figures 4 and 5, in some implementations, the back-diffusion process based on the diffusion model successively denoises the training image of the resulting noise state through a denoising network to train multiple denoising networks (i.e., 023), including:

[0072] 0231: Use the training image of the resulting noise state as the training image of the current noise state;

[0073] 0232: Input the training image of the current noise state and the corresponding time series into the denoising network of the corresponding state transition interval, so as to output the noise to be eliminated corresponding to the denoising network;

[0074] 0233: Eliminate the noise to be eliminated from the training image of the current noise state to obtain the training image of the next noise state, and use it again as the training image of the current noise state. Then return to the step of inputting the training image of the current noise state and the corresponding time series into the denoising network of the corresponding state transition interval to output the noise to be eliminated corresponding to the denoising network.

[0075] 0234: Based on each noise to be eliminated and the corresponding random Gaussian noise, a corresponding constraint relationship is established through the mean absolute value error loss;

[0076] 0235: Train multiple denoising networks based on multiple constraints.

[0077] In some implementations, the result noise state The training image is used as the training image for the current noise state, and the training image for the current noise state and the corresponding time series are input into the denoising network for the corresponding state transition interval. The current noise state is the resulting noise state. At that time, the corresponding time series is The corresponding state transition interval is also the th State transition interval, the first The denoising network for the state transition interval is represented as follows .

[0078] Through denoising network Operations such as downsampling and upsampling coding in the code can predict and output the noise to be eliminated in this state transition interval, denoted as... Remove the noise to be removed from the training image in the current noise state. This allows us to obtain the training image for the next noisy state, i.e., the first... Noise status The training images.

[0079] In the forward diffusion process of the diffusion model, from the first... Noise status To the result noise state The added random Gaussian noise is Based on the noise to be eliminated With the corresponding random Gaussian noise A corresponding constraint relationship can be established using the Mean Absolute Error (MAE) loss. MAE loss is also known as L1 loss, the first... The constraints of the state transition interval are shown below:

[0080]

[0081] Through the first The constraints on the state transition interval can be applied to the first... Denoising network for state transition intervals Training is performed to make the denoising network It is possible to predict more accurate noise to be eliminated, thereby enabling the first Noise status The training images and the first Noise status The training images should be as close as possible to each other or as consistent as possible.

[0082] After obtaining the number Noise status After training the images, the first one can be... Noise status The training image is re-used as the training image for the current noise state. Then, the training image for the current noise state and the corresponding time series are input into the denoising network for the corresponding state transition interval to output the step of the noise to be eliminated corresponding to the denoising network.

[0083] And so on, until the first... Noise status After training the images, the first one can be... Noise status The training image is re-used as the training image for the current noise state. Then, the training image for the current noise state and the corresponding time series are input into the denoising network for the corresponding state transition interval to output the step of the noise to be eliminated corresponding to the denoising network.

[0084] In the current noise state, the number is... Noise status At that time, the corresponding time series is The corresponding state transition interval is also the th State transition interval, the first The denoising network for the state transition interval is represented as follows .

[0085] Through denoising network Operations such as downsampling and upsampling coding in the code can predict and output the noise to be eliminated in this state transition interval, denoted as... Remove the noise to be removed from the training image in the current noise state. This allows us to obtain the training image for the next noisy state, i.e., the first... Noise status The training images.

[0086] In the forward diffusion process of the diffusion model, from the initial noise state To the first noise state The added random Gaussian noise is Based on the noise to be eliminated With the corresponding random Gaussian noise A corresponding constraint relationship can be established using MAE loss. The constraints of the state transition interval are shown below:

[0087]

[0088] Through the first The constraints on the state transition interval can be applied to the first... Denoising network for state transition intervals Training is performed to make the denoising network It is possible to predict more accurate noise to be eliminated, thereby enabling the first Noise status Training images and initial noise state The training images should be as close or identical as possible.

[0089] In this way, by training multiple denoising networks, it is possible to achieve the goal of making the training image of the noise state obtained after eliminating the noise to be eliminated predicted by each denoising network as close as possible or even consistent with the training image of the initial noise state during the back-diffusion process. Thus, multiple denoising networks have the ability to simulate and generate high-quality palmprint feature images from random noise.

[0090] After multiple denoising networks have been trained, multiple initial feature images can be generated based on the simulation of random noise batches by the multiple denoising networks.

[0091] Referring to Figure 6, in some embodiments, before performing intra-class augmentation on multiple initial palm feature images to obtain the target palm feature image (i.e., 040), the method for simulating and generating palm feature data further includes:

[0092] 050: Calculate the image similarity and texture distance between every two initial palm feature images;

[0093] 060: When the image similarity is greater than the preset similarity or the texture distance is less than the preset distance, the two initial palm feature images are determined to be similar, and one initial palm feature image is removed from the two initial palm feature images;

[0094] 070: When the image similarity is less than or equal to the preset similarity and / or the texture distance is greater than or equal to the preset distance, the two initial palm feature images are determined to be dissimilar, and the two initial palm feature images are retained.

[0095] In some implementations, considering that there may be cases where the texture similarity among the multiple initial palm feature images generated is too high, the initial palm feature images that are too similar can be filtered out before performing intra-class expansion on the multiple initial palm feature images.

[0096] First, calculate the image similarity and texture distance between every two initial palm feature images. Structural similarity (SSIM) can be used to represent image similarity; the image similarity between two initial palm feature images can be expressed as... The value of SSIM ranges from [0, 1]. The larger the value of SSIM, the more similar the two initial palm feature images are.

[0097] Texture distance can be represented using Deep Image Structure and Texture Similarity (DISTS). The texture distance between two initial hand feature images can be expressed as: The value of DISTS is greater than or equal to zero. A DISTS value of zero indicates that the two initial palm feature images are completely identical. The closer the DISTS value is to zero, the more similar the two initial palm feature images are.

[0098] Set judgment thresholds for image similarity and texture distance respectively. The judgment threshold for image similarity is a preset similarity. Since the value range of SSIM is [0, 1], the range of the preset similarity is also [0, 1]. The preset similarity can be set according to the actual application situation and is not restricted here. For example, the preset similarity can be set to any value such as 0.9, 0.95, or 0.98. If the image similarity is greater than the preset similarity, it means that the two initial palm feature images are relatively similar.

[0099] The threshold for determining texture distance is a preset distance. This preset distance can be set according to the actual application and is not limited here; for example, it can be set to any value such as 0.1, 0.3, or 0.5. If the texture distance is less than the preset distance, it indicates that the two initial hand feature images are quite similar.

[0100] The algorithm combines image similarity and texture distance to determine the similarity between any two initial hand feature images. The calculated image similarity is compared to a preset similarity, and the calculated texture distance is compared to a preset distance. If the image similarity is greater than the preset similarity or the texture distance is less than the preset distance, the two initial hand feature images are considered similar and belong to the same category. In this case, one initial hand feature image is randomly removed from the two initial hand feature images.

[0101] If the image similarity is less than or equal to a preset similarity and / or the texture distance is greater than or equal to a preset distance, the two initial palm feature images are determined to be dissimilar and can be classified as two independent images. In this case, the two initial palm feature images can be retained.

[0102] In this way, images with overly similar textures among multiple initial palm feature images can be filtered out, ensuring that the inter-class distances of different initial palm feature images are reasonable. This ensures that the final generated palm feature data can stably support the performance improvement of the palm feature recognition model.

[0103] After filtering out overly similar initial hand feature images from multiple initial hand feature images, the remaining initial hand feature images can be classified into separate categories for intra-class expansion. The specific process of intra-class expansion is described below.

[0104] Referring to Figures 7 and 8, in some embodiments, the target palm feature image includes an initial palm feature image, a first target image, a second target image, and a third target image. Intra-class expansion is performed on multiple initial palm feature images to obtain multiple target palm feature images (i.e., 040), including:

[0105] 041: Select multiple basic operations from the basic operation library and arrange them in a pluralistic way to obtain multiple operation combinations;

[0106] 042: Process each initial palm feature image according to multiple operation combinations to obtain multiple first target images corresponding to each initial palm feature image;

[0107] 043: Determine the region to be processed in the initial palm feature image;

[0108] 044: Select multiple enhancement operations from the enhancement operation library to process the region to be processed in each initial palm feature image, and obtain multiple second target images corresponding to each initial palm feature image;

[0109] 045: Select multiple simulation operations from the simulation operation library to process each initial palm feature image, and obtain multiple third target images corresponding to each initial palm feature image.

[0110] In some implementations, intra-class expansion can be divided into three methods.

[0111] The first expansion method involves setting up a basic operation library. This library includes multiple basic operations, and the specific operations and their number can be configured according to actual application needs. In one example, the basic operation library includes any combination of sharpening, contrast adjustment, translation, and rotation operations, as shown in Figure 8. Of course, the basic operation library can also include any other basic image processing operations.

[0112] Select some or all of the basic operations from the basic operation library and arrange them into multiple operation combinations. Then, process each initial palm feature image according to each of these multiple operation combinations. Processing each initial palm feature image according to each operation combination yields a corresponding first target image; processing each initial palm feature image according to multiple operation combinations yields multiple corresponding first target images. The number of first target images is equal to the number of operation combinations.

[0113] For example, five basic operations can be selected and arranged in five arbitrary combinations to obtain five operation combinations. Each initial palm feature image is then processed according to the five operation combinations, resulting in five corresponding first target images for each initial palm feature image.

[0114] Multiple first target images corresponding to each initial palm feature image obtained through the first expansion method can simulate the influence of basic environmental factors such as illumination, translation, and rotation on the palm features of the same palm.

[0115] The second expansion method involves defining a region to be processed within each initial palm feature image. This region can be rectangular. An enhancement operation library is then established, containing multiple enhancement operations. The specific operations and their number can be configured according to actual application requirements. In one example, the library includes any combination of Gaussian blur, pixel loss, and edge overlay operations. For instance, as shown in Figure 8, the library includes Gaussian blur, pixel loss, and edge overlay operations. Of course, the library can also include any other arbitrary image enhancement operations.

[0116] Multiple enhancement operations are selected from the enhancement operation library to process the region to be processed in each initial palm feature image. The selected enhancement operations can be all or some of the enhancement operations in the library. Processing the region to be processed in each initial palm feature image according to each enhancement operation yields a corresponding second target image; processing the region to be processed in each initial palm feature image according to multiple enhancement operations yields multiple corresponding second target images. The number of second target images is equal to the number of selected enhancement operations.

[0117] For example, two enhancement operations can be selected from the enhancement operation library to process the region to be processed in each initial palm feature image, resulting in two corresponding second target images for each initial palm feature image. The enhancement operations only process the region to be processed; the regions in the initial palm feature image outside the region to be processed remain unchanged.

[0118] Multiple second target images corresponding to each initial palm feature image obtained through the second expansion method can simulate the changes in local palm features of the same palm under slight expansion or contraction.

[0119] The third expansion method involves setting up a simulation operation library. This library includes multiple simulation operations, and the specific operations and their number can be configured according to actual application requirements. In one example, the simulation operation library includes any combination of operations such as motion blur, projection transformation, affine transformation, global distortion, and local distortion, as shown in Figure 8. Of course, the simulation operation library can also include any other arbitrary image transformation operations.

[0120] Multiple simulation operations are selected from the simulation operation library to process each initial palm feature image. The selected simulation operations can be all or some of the simulation operations in the library. Processing each initial palm feature image according to each simulation operation yields a corresponding third target image; processing each initial palm feature image according to multiple simulation operations yields multiple corresponding third target images. The number of third target images is equal to the number of selected simulation operations.

[0121] For example, two simulation operations can be selected from the simulation operation library to process each initial palm feature image, and each initial palm feature image can yield two corresponding third target images.

[0122] In the third extension method, motion simulation can simulate the blurring effect of hand features caused by hand movement under real-world application conditions; projection transformation and affine transformation operations can simulate the relative positional shift of hand features caused by hand tilting under real-world application conditions. Global distortion and local distortion operations can simulate the changes in hand features caused by hand extension under real-world application conditions.

[0123] It should be noted that sharpening, contrast adjustment, X-axis translation, Y-axis translation, rotation, Gaussian blur, pixel loss, edge overlay, motion blur, projection transformation, affine transformation, global distortion, and local distortion can all be performed using conventional methods and logic, so they will not be elaborated here.

[0124] In this embodiment, multiple initial palm feature images are augmented within a class using three different augmentation methods to obtain multiple target palm feature images. This approach considers the impact of environmental factors, palm movement and tilt, and other real-world conditions on palm feature recognition. By simulating the influence of these factors on palm features through intra-class augmentation, the robustness of the simulated palm feature data is increased, and the intra-class diversity of the simulated palm feature data is ensured, thereby improving the robustness of the palm feature recognition model.

[0125] Referring to Figure 9, this application also provides a palm feature data simulation generation device 100. The simulation generation device 100 includes a construction module 10, a training module 20, a generation module 30, and an expansion module 40. The construction module 10 is used to construct a palm feature learning database. The training module 20 is used to train multiple denoising networks based on the palm feature learning database. The generation module 30 is used to batch generate multiple initial palm feature images based on the trained multiple denoising networks. The expansion module 40 is used to perform intra-class expansion on the multiple initial palm feature images to obtain multiple target palm feature images.

[0126] In some implementations, the palm feature learning database includes feature learning images. The construction module 10 is specifically used to acquire a real, complete palm image; perform keypoint detection on the complete palm image; perform image angle correction on the complete palm image based on the detected keypoints to obtain a corrected complete palm image; and crop the region of interest from the corrected complete palm image to obtain the feature learning image.

[0127] In some implementations, the training module 20 is specifically used to determine training images from a palm feature learning database; based on the forward diffusion process of the diffusion model, random Gaussian noise is added to the training images one by one, so that the training images change from the initial noise state to the result noise state; based on the backward diffusion process of the diffusion model, the training images in the result noise state are denoised one by one through a denoising network to train multiple denoising networks.

[0128] In some implementations, the training module 20 is specifically used to: use the training image of the resulting noise state as the training image of the current noise state; input the training image of the current noise state and the corresponding time series into the denoising network of the corresponding state transition interval to output the noise to be eliminated corresponding to the denoising network; eliminate the noise to be eliminated from the training image of the current noise state to obtain the training image of the next noise state, and use it again as the training image of the current noise state; then return to the step of inputting the training image of the current noise state and the corresponding time series into the denoising network of the corresponding state transition interval to output the noise to be eliminated corresponding to the denoising network; establish a corresponding constraint relationship based on each noise to be eliminated and the corresponding random Gaussian noise through the mean absolute value error loss; and train multiple denoising networks according to multiple constraint relationships.

[0129] Referring to Figure 10, in some embodiments, the simulation generation device 100 further includes a judgment module 50. Before performing intra-class expansion on multiple initial palm feature images to obtain a target palm feature image, the judgment module 50 calculates the image similarity and texture distance between every two initial palm feature images; when the image similarity is greater than a preset similarity or the texture distance is less than a preset distance, the two initial palm feature images are determined to be similar, and one initial palm feature image is removed from the two initial palm feature images; when the image similarity is less than or equal to a preset similarity and / or the texture distance is greater than or equal to a preset distance, the two initial palm feature images are determined to be dissimilar, and the two initial palm feature images are retained.

[0130] In some embodiments, the target palm feature image includes an initial palm feature image, a first target image, a second target image, and a third target image. The expansion module 40 is specifically configured to select multiple basic operations from a basic operation library and arrange them in a pluralistic manner to obtain multiple operation combinations; process each initial palm feature image according to the multiple operation combinations to obtain multiple first target images corresponding to each initial palm feature image; determine the region to be processed in the initial palm feature image; select multiple enhancement operations from an enhancement operation library to process the region to be processed in each initial palm feature image to obtain multiple second target images corresponding to each initial palm feature image; and select multiple simulation operations from a simulation operation library to process each initial palm feature image to obtain multiple third target images corresponding to each initial palm feature image.

[0131] It should be noted that the explanation of the palm feature data simulation generation method in the foregoing embodiments also applies to the palm feature data simulation generation device 100 in the embodiments of this application, and will not be elaborated here.

[0132] Referring to Figure 11, this application also provides a palm feature data simulation generation system 200. The palm feature data simulation generation system 200 includes one or more processors 210 and a memory 220. The memory 220 stores a computer program. When the computer program is executed by the processor 210, it implements the palm feature data simulation generation method of any of the above embodiments.

[0133] For example, when the computer program is executed by the processor 210, the following method for simulating and generating palm feature data is implemented:

[0134] 010: Construct a palm feature learning database;

[0135] 020: Training multiple denoising networks based on a palm feature learning database;

[0136] 030: Generate multiple initial palm feature images in batches based on multiple trained denoising networks;

[0137] 040: Perform intra-class expansion on multiple initial palm feature images to obtain multiple target palm feature images.

[0138] For example, when the computer program is executed by the processor 210, the following method for simulating and generating palm feature data is implemented:

[0139] 011: Obtain a realistic, complete image of the palm;

[0140] 012: Perform key point detection on the complete palm image;

[0141] 013: Based on the detected key points, perform image angle correction on the complete palm image to obtain the corrected complete palm image;

[0142] 014: Cropping the region of interest from the corrected complete hand image yields the feature-learned image.

[0143] It should be noted that the explanation of the palm feature data simulation generation method and simulation generation device 100 in the foregoing embodiments also applies to the palm feature data simulation generation system 200 in the embodiments of this application, and will not be elaborated here.

[0144] Please refer to Figure 12. This application also provides a computer-readable storage medium 300 storing a computer program 310, which, when executed by a processor 320, implements a method for simulating and generating palm feature data according to any of the above embodiments.

[0145] For example, when computer program 310 is executed by processor 320, the following method for simulating and generating palm feature data is implemented:

[0146] 010: Construct a palm feature learning database;

[0147] 020: Training multiple denoising networks based on a palm feature learning database;

[0148] 030: Generate multiple initial palm feature images in batches based on multiple trained denoising networks;

[0149] 040: Perform intra-class expansion on multiple initial palm feature images to obtain multiple target palm feature images.

[0150] For example, when computer program 310 is executed by processor 320, the following method for simulating and generating palm feature data is implemented:

[0151] 011: Obtain a realistic, complete image of the palm;

[0152] 012: Perform key point detection on the complete palm image;

[0153] 013: Based on the detected key points, perform image angle correction on the complete palm image to obtain the corrected complete palm image;

[0154] 014: Cropping the region of interest from the corrected complete hand image yields the feature-learned image.

[0155] It should be noted that the explanation of the simulation generation method and simulation generation device 100 for palm feature data in the foregoing embodiments also applies to the computer-readable storage medium 300 of the embodiments of this application, and will not be elaborated here.

[0156] In summary, the palm feature data simulation generation method, simulation generation apparatus 100, simulation generation system 200, and computer-readable storage medium 300 of this application train multiple denoising networks based on a palm feature learning database, and generate multiple initial palm feature images in batches based on the trained multiple denoising networks. Then, intra-class expansion is performed on the multiple initial palm feature images to obtain multiple target palm feature images. In this way, high-quality and highly diverse palm feature data can be generated in batches, meeting the training requirements of palm feature recognition models and improving the robustness of palm recognition models.

[0157] For example, in terms of improving R&D efficiency, there is a high demand for palm recognition products in industrial applications such as smart access control. However, the cost of collecting real data is high and the sample size is limited, which affects the R&D progress. This method can provide sufficient data quickly and at low cost, shorten the R&D cycle, and bring products to market faster.

[0158] Regarding enhancing model robustness, the complex industrial environment, with factors such as lighting, pose, and occlusion causing variations in the distribution of hand feature data, can easily lead to recognition errors. The target hand feature images generated by this method cover a variety of complex situations, allowing the model to learn more comprehensive features during training, thus enhancing its adaptability to different industrial scenarios and improving recognition accuracy.

[0159] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, different embodiments or examples described in this specification, as well as features of different embodiments or examples, may be combined and integrated.

[0160] Any process or method description in the flowchart or otherwise described herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved.

[0161] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, a computer-readable storage medium can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable storage medium could be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0162] The various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0163] The implementation of all or part of the steps carried out by the methods in the above embodiments can be accomplished by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes one or a combination of the steps of the method embodiments. Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The aforementioned storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0164] Although embodiments of this application have been shown and described above, the above embodiments are exemplary and should not be construed as limiting this application. Variations, modifications, substitutions and modifications can be made to the above embodiments within the scope of this application, which is defined by the claims and their equivalents.

Claims

1. A method for simulating and generating palm feature data, wherein, include: Construct a palm feature learning database; Multiple denoising networks are trained based on the aforementioned palm feature learning database; Multiple initial palm feature images are generated in batches based on the trained denoising networks. Multiple initial palm feature images are expanded within a class to obtain multiple target palm feature images.

2. The method for simulating and generating palm feature data according to claim 1, wherein, The palm feature learning database includes feature learning images, and constructing the palm feature learning database includes: Obtain a realistic, complete image of the palm; Key point detection is performed on the complete palm image; Based on the detected key points, the image angle of the complete palm image is corrected to obtain the corrected complete palm image; The region of interest is cropped from the corrected complete hand image to obtain the feature-learned image.

3. The method for simulating and generating palm feature data according to claim 1, wherein, The training of multiple denoising networks based on the palm feature learning database includes: Training images are determined from the palm feature learning database; Based on the forward diffusion process of the diffusion model, random Gaussian noise is added to the training image one by one, so that the training image changes from the initial noise state to the result noise state. Based on the back-diffusion process of the diffusion model, the training image in the result noise state is successively denoised through the denoising network to train multiple denoising networks.

4. The method for simulating and generating palm feature data according to claim 3, wherein, The back-diffusion process based on the diffusion model successively denoises the training image in the resulting noise state through the denoising network to train multiple denoising networks, including: The training image of the resulting noise state is used as the training image of the current noise state; The training image and its corresponding time series in the current noise state are input into the denoising network in the corresponding state transition interval to output the noise to be eliminated corresponding to the denoising network. The training image of the current noise state is used to remove the noise to be removed, and the training image of the next noise state is obtained. This image is then used as the training image of the current noise state again. The process is then repeated, and the process is then repeated, where the training image of the current noise state and the corresponding time series are input into the denoising network of the corresponding state transition interval to output the noise to be removed corresponding to the denoising network. Based on each of the noises to be eliminated and the corresponding random Gaussian noise, a corresponding constraint relationship is established through the average absolute value error loss. The denoising networks are trained according to the various constraints.

5. The method for simulating and generating palm feature data according to claim 1, wherein, Before performing intra-class augmentation on the multiple initial palm feature images to obtain the target palm feature image, the method for simulating and generating the palm feature data further includes: Calculate the image similarity and texture distance between every two initial palm feature images; When the image similarity is greater than a preset similarity or the texture distance is less than a preset distance, the two initial palm feature images are determined to be similar, and one of the initial palm feature images is removed from the two initial palm feature images; When the image similarity is less than or equal to the preset similarity and / or the texture distance is greater than or equal to the preset distance, the two initial palm feature images are determined to be dissimilar, and the two initial palm feature images are retained.

6. The method for simulating and generating palm feature data according to claim 1, wherein, The target palm feature image includes the initial palm feature image, the first target image, the second target image, and the third target image. The process of expanding the multiple initial palm feature images within a class to obtain multiple target palm feature images includes: Multiple basic operations are selected from the basic operation library and arranged and combined to obtain multiple operation combinations; Each initial palm feature image is processed according to a plurality of the operation combinations to obtain a plurality of first target images corresponding to each initial palm feature image; The region to be processed is determined in the initial palm feature image; Multiple enhancement operations are selected from the enhancement operation library to process the region to be processed in each initial palm feature image, thereby obtaining multiple second target images corresponding to each initial palm feature image; Multiple simulation operations are selected from the simulation operation library to process each of the initial palm feature images, thereby obtaining multiple third target images corresponding to each initial palm feature image.

7. The method for simulating and generating palm feature data according to claim 6, wherein, The basic operation library includes any combination of operations such as sharpening, contrast adjustment, panning, and rotation; and / or The enhanced operation library includes any combination of operations such as Gaussian blur, pixel loss, and edge overlay; and / or The simulation operation library includes any number of operations such as motion blur, projection transformation, affine transformation, global distortion, and local distortion.

8. A device for simulating and generating palm feature data, wherein, include: Build modules are used to construct a palm feature learning database; The training module is used to train multiple denoising networks based on the palm feature learning database; The generation module is used to generate multiple initial palm feature images in batches based on the multiple trained denoising networks; An expansion module is used to expand multiple initial palm feature images within a class to obtain multiple target palm feature images.

9. A simulation generation system for palm feature data, wherein, The palm feature data simulation generation system includes one or more processors and a memory, the memory storing a computer program, which, when executed by the processor, implements the palm feature data simulation generation method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, wherein, When the program is executed by a processor, it implements the method for simulating and generating palm feature data as described in any one of claims 1-7.