WO2025103010A1 - 掌纹图像生成方法、装置、设备、存储介质及程序产品 - Google Patents

掌纹图像生成方法、装置、设备、存储介质及程序产品 Download PDF

Info

Publication number
WO2025103010A1
WO2025103010A1 PCT/CN2024/123335 CN2024123335W WO2025103010A1 WO 2025103010 A1 WO2025103010 A1 WO 2025103010A1 CN 2024123335 W CN2024123335 W CN 2024123335W WO 2025103010 A1 WO2025103010 A1 WO 2025103010A1
Authority
WO
WIPO (PCT)
Prior art keywords
palmprint
line
palm
map
line energy
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/123335
Other languages
English (en)
French (fr)
Inventor
沈雷
金建龙
张睿欣
张菁芸
丁守鸿
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Publication of WO2025103010A1 publication Critical patent/WO2025103010A1/zh
Anticipated expiration legal-status Critical
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—Two-dimensional [2D] image generation
    • G06T11/10—Texturing; Colouring; Generation of textures or colours
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/12—Fingerprints or palmprints
    • G06V40/1347—Preprocessing; Feature extraction
    • G06V40/1359—Extracting features related to ridge properties; Determining the fingerprint type, e.g. whorl or loop
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/12—Fingerprints or palmprints
    • G06V40/1365—Matching; Classification

Definitions

  • the present application relates to the field of artificial intelligence, and more specifically, to a palmprint image generation method, device, equipment, storage medium and program product.
  • Palmprints refer to the skin lines inside the palm of a person, which are unique and stable and can be used for individual identification and identity verification.
  • the acquisition and annotation of palmprint data are relatively complex and time-consuming, resulting in a limited number of available palmprint datasets, especially large-scale datasets.
  • Palmprints are part of personal body features and are sensitive and private. Therefore, when collecting and using palmprint data, we must comply with relevant privacy laws and regulations and take appropriate security measures to protect user privacy.
  • the present application introduces a new palm line energy (Palm Crease Energy, PCE) domain, first converts the Bezier curve into the palm line energy domain to generate a palm line energy map with realistic wrinkles, and then generates a realistic palm print image with realistic texture based on the palm line energy map, thereby generating diversified and realistic palm print images.
  • PCE Palm Line Energy
  • the embodiments of the present application provide a palmprint image generation method, apparatus, device, storage medium and program product.
  • an embodiment of the present application provides a palmprint image generation method, which is executed by an electronic device and includes:
  • a predetermined palmprint curve template Based on a predetermined palmprint curve template, a plurality of control points are determined, and a Bezier curve is generated based on the plurality of control points;
  • a palm line energy map with wrinkle information is generated, wherein the palm line energy map includes palm lines having the same line distribution as the Bezier curve but different line types, wherein the line type is used to describe the wrinkle information and corresponds to the line direction energy of each pixel on the palm line; and,
  • a simulated palmprint image with detailed texture information is generated.
  • an embodiment of the present application provides a palmprint image generating device, comprising:
  • a curve generating module is configured to determine a plurality of control points based on a predetermined palmprint curve template, and generate a Bezier curve based on the plurality of control points;
  • a wrinkle generation module is configured to generate a palm line energy map with wrinkle information based on the Bezier curve, wherein the palm line energy map includes palm lines having the same line distribution as the Bezier curve but different line types, wherein the line type is used to describe the wrinkle information and corresponds to the line direction energy of each pixel on the palm line;
  • the texture generation module is configured to generate a simulated palm print image with detailed texture information based on the palm print line energy map.
  • an embodiment of the present application provides an electronic device, comprising: one or more processors; and one or more memories, wherein a computer executable program is stored in the one or more memories, and when the computer executable program is executed by the processor, the palmprint image generation method as described above is executed.
  • an embodiment of the present application provides a computer-readable storage medium having computer-executable instructions stored thereon, which are used to implement the palmprint image generation method as described above when executed by a processor.
  • the embodiment of the present application provides a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the palmprint image generation method according to the embodiment of the present application.
  • FIG1 is a schematic diagram showing the significant difference between the Bezier curve and the real palm print in terms of wrinkle distribution and texture
  • FIG2 is a flow chart showing a palmprint image generating method according to an embodiment of the present application.
  • FIG3 is a schematic diagram showing a palmprint image generating system according to an embodiment of the present application.
  • FIG4A is a schematic diagram showing an example palmprint curve template according to an embodiment of the present application.
  • FIG4B is a schematic diagram showing a comparison between a Gaussian-MFRAT kernel according to an embodiment of the present application and a traditional MFRAT filter;
  • FIG5 is a schematic diagram illustrating domain conversion in joint training according to an embodiment of the present application.
  • FIG6 is a schematic diagram showing a first generation phase in joint training according to an embodiment of the present application.
  • FIG7 is a schematic diagram showing a palm line energy extractor according to an embodiment of the present application.
  • FIG8 is a schematic diagram showing a second generation phase in joint training according to an embodiment of the present application.
  • FIG9 is a comparison diagram showing palmprint generation results using different palmprint generation methods according to an embodiment of the present application.
  • FIG10 is a diagram showing the result of the effectiveness verification of the line energy feature enhancement block according to an embodiment of the present application.
  • FIG11 is a schematic diagram showing a palmprint image generating device according to an embodiment of the present application.
  • FIG12 shows a schematic diagram of an electronic device according to an embodiment of the present application.
  • FIG. 13 is a schematic diagram showing the architecture of an exemplary computing device according to an embodiment of the present application.
  • the palm print image generation method of the present application can be implemented based on artificial intelligence (AI).
  • Artificial intelligence is the use of digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, so that the machine has the functions of perception, reasoning and decision-making.
  • artificial intelligence studies the design principles and implementation methods of various intelligent machines, so that the palm print image generation method of the present application can achieve the following functions: based on the control points determined from the palm print curve template, a realistic palm print image with realistic wrinkles and texture is generated.
  • the palmprint image generation method of the present application can also be implemented based on computer vision (CV) technology.
  • Computer vision technology can obtain information from images or multi-dimensional data.
  • the palmprint image generation method of the present application can use CV technology to generate a palmprint line energy map with a style close to the real palmprint from the Bezier curve, and then generate a palmprint image with diversified texture information based on the palmprint line energy map to achieve diversified palmprint image output, which is used for pre-training processes such as palmprint recognition models.
  • the palm print image generation method of the present application can be implemented based on Bezier curve.
  • Bezier curve is used to describe a smooth curve.
  • Bezier curve can be applied to the drawing of palm print line.
  • the characteristic of Bezier curve is that the shape of the curve can be controlled by control points, so the curvature of the palm print line and the shape of the curve can be adjusted by control points.
  • the palm can be divided into several sections, each section is described by a Bezier curve, and palm print lines of different shapes can be obtained by adjusting the position and number of control points.
  • FIG. 1 is a schematic diagram showing that there is a significant gap between the Bezier curve and the real palm print in terms of wrinkle distribution and texture.
  • Palmprint as a stable and privacy-friendly biometric recognition technology, has recently shown great potential in recognition applications.
  • deep learning-based palmprint recognition methods have become the mainstream palmprint recognition technology.
  • Deep learning-based palmprint recognition methods train neural networks to extract features of palmprints with improved classification or pairing losses.
  • a major difficulty in the research and application of deep learning-based palmprint recognition is the scarcity of large-scale palmprint datasets, and collecting large-scale palmprint datasets may pose the risk of violating user privacy.
  • researchers can currently expand the dataset by using some data synthesis techniques to generate simulated palmprint data.
  • the Bezier palmprint generation method uses parameterized Bezier curves to synthesize pseudo palmprint lines.
  • the Bezier palmprint 101 generated by the Bezier palmprint generation method has significant differences in wrinkles and textures from the real palmprint image 102. It cannot reflect the wrinkle distribution of the real palmprint, nor can it present the various detailed textures in the real palmprint. Therefore, the Bezier palmprint still requires a certain amount of real palmprint data for fine-tuning.
  • GANs generative adversarial networks
  • the embodiment of the present application provides a method of using an intermediate domain connecting the Bezier palmprint domain and the real palmprint image domain, wherein a new palmprint line energy (PCE) domain is introduced as the intermediate domain.
  • PCE palmprint line energy
  • the Bezier curve is converted to the palmprint line energy domain to generate a palmprint line energy map with realistic wrinkles (referred to as PCE image), and then based on the palmprint line energy map, a simulated palmprint image with realistic texture is generated, thereby obtaining a diversified and realistic palmprint image.
  • the PCE image 103 is an intermediate state close to the real palm print. It has wrinkle consistency with the Bezier curve 101, that is, the line distribution of its wrinkles (i.e., palm print lines) is consistent with the Bezier curve 101, and has appearance similarity with the real palm print image 102, that is, the appearance of the palm print lines of the PCE image 103 is consistent with the appearance of the real palm print lines.
  • the method provided in the embodiment of the present application decomposes the Bezier-Real difference into the wrinkle difference and the texture difference, thereby reducing the generation difficulty.
  • the generation of palmprint wrinkles and textures is decoupled, so as to generate realistic wrinkles by converting the Bezier curve in the Bezier palmprint domain to the palmprint line energy map in the palmprint line energy domain, and to generate realistic textures by converting the palmprint line energy map in the palmprint line energy domain to the palmprint image in the palmprint image domain, thereby reducing the difficulty of generating realistic palmprint images from Bezier curves, and realizing diversified realistic palmprint image generation.
  • the method provided in the embodiment of the present application generates a Bezier curve using control points determined from a predetermined palmprint curve template, and converts the Bezier curve into a palmprint line energy map with wrinkle information, wherein the wrinkle information includes the same line distribution as the Bezier curve but a different line type, and the line type is determined by the line energy feature of each pixel, and then further generates a palmprint image with texture information based on the palmprint line energy map with the wrinkle information, and the palmprint lines of the palmprint image are consistent with the wrinkle information of the palmprint line energy map, thereby generating a simulated palmprint image with realistic wrinkles and texture.
  • the method provided in the embodiment of the present application introduces a palmprint line energy domain as an intermediate domain connecting the Bezier palmprint domain and the palmprint image domain, thereby avoiding directly generating a palmprint image with wrinkle information and texture information from a Bezier curve, thereby reducing the difficulty of generating a palmprint image.
  • a palmprint line energy domain as an intermediate domain connecting the Bezier palmprint domain and the palmprint image domain, thereby avoiding directly generating a palmprint image with wrinkle information and texture information from a Bezier curve, thereby reducing the difficulty of generating a palmprint image.
  • detailed texture information is generated while ensuring that the palmprint image has consistent palmprint lines, so that a realistic palmprint image with diversified textures can be generated while retaining the same identity information, thereby reducing dependence on real data and being suitable for palmprint recognition training in the absence of a large-scale palmprint data set.
  • Fig. 2 is a flow chart showing a palmprint image generation method 200 according to an embodiment of the present application.
  • Fig. 3 is a schematic diagram showing a palmprint image generation system according to an embodiment of the present application.
  • the palmprint image generation system may include a first generation stage 310 associated with a first generator 304, and a second generation stage 320 associated with a second generator 307.
  • the first generation stage 310 and the second generation stage 320 are connected by a PCE image 305 of the PCE domain.
  • the Bezier palmprint generator 302 In the first generation stage 310 , the Bezier palmprint generator 302 generates a Bezier curve 303 based on the control point 301 , and the first generator 304 generates a PCE image 305 based on the Bezier curve 303 .
  • the second generator 307 In the second generation stage 320, the second generator 307 generates a simulated palm print image 308 based on the PCE image 305 and the control vector 306. The generation process will be described in detail below with reference to FIG. 2 to FIG. 4A.
  • a plurality of control points may be determined based on a predetermined palmprint curve template, and a Bezier curve may be generated based on the plurality of control points.
  • the palmprint image generation method can use prior knowledge obtained from human skin texture to improve the control point generation mechanism, for example, to improve the palmprint curve template used to generate control points.
  • the palm print curve template can be predetermined based on the statistical information obtained from the real palm print lines of humans. Since the diversity and individual differences of real human palm prints are taken into consideration at the same time, the palm print curve template determined is more representative. Therefore, the generation range of the control points determined based on the palm print curve template will be more accurate, so that a Bezier curve that is closer to the distribution of real palm print wrinkles can be generated.
  • FIG. 4A is a schematic diagram showing an example palm print curve template according to an embodiment of the present application.
  • five example palm print curve templates divided based on statistical information obtained from real human palm print lines are provided, corresponding to the five most representative human palm print line distributions.
  • each example palm print curve template provides a generation range for control points, as shown in the dotted box.
  • Different example palm print curve templates may have different numbers of palm print lines, and therefore may be determined by different numbers of generation ranges and control points.
  • example palmprint curve template in FIG. 4A is not intended to limit the palmprint curve template used to generate control points in the present application, and the present application may also adopt various other palmprint curve templates.
  • determining a plurality of control points may include:
  • Sampling is performed within the region to obtain the plurality of control points.
  • the sampling method may include methods such as random sampling, so as to increase a certain degree of randomness and individual differences while maintaining the shape and style of the palm print, so as to make the generated palm print image more realistic and diverse.
  • Other sampling methods may also be used to achieve different data generation effects, and this application does not limit this.
  • these palm print curve templates can provide a more accurate generation range of control points, so that the Bezier curves generated based on these control points are closer to the distribution of real wrinkles, thereby reducing the difference between the generated palm prints and the real palm prints.
  • a palm line energy map with wrinkle information can be generated based on the Bezier curve, and the palm line energy map includes palm lines with the same line distribution as the Bezier curve but different line types, wherein the line type is used to describe the wrinkle information and corresponds to the line direction energy of each pixel on the palm line.
  • the Bezier curve 303 can be converted to the PCE domain to generate a PCE image (i.e., palm line energy map) 305.
  • a PCE image i.e., palm line energy map
  • Both the Bezier curve 303 and the PCE image 305 are binary images based on lines.
  • the goal of the first generation stage is to convert the curve lines in the Bezier palm print image into a PCE image with wrinkles (i.e., palm lines) that are closer to real human palm prints.
  • wrinkles can refer to the deep and shallow concave and convex lines in the palm lines, which are usually formed due to the folding and bending of the skin. They are used to describe the overall shape of the palm lines, such as the main lines of the palm, curved edges, etc., and they play a role in segmenting and defining different areas in the palm lines.
  • the shape and distribution of palm lines can be used in fields such as individual identity recognition.
  • the PCE image can include palm lines with the same line distribution (i.e., the position and direction of the main lines of the palm are consistent), but the palm lines in the PCE image have different line types (e.g., lines of different thickness and depth) to simulate the wrinkles of real human palm lines.
  • the different line types used on the palm line may depend on the line energy characteristics of the pixels at the corresponding positions on the palm line, wherein the line energy characteristics of each pixel may be obtained by enhancing the line energy characteristics of the Bezier curve, which may describe the directional distribution of the line energy characteristics at the pixel.
  • generating a palm line energy map based on the Bezier curve may include:
  • the palm line energy map is generated based on the enhanced multi-channel feature map.
  • a Bezier curve has N channels.
  • the line energy feature enhancement can be performed on the feature map on each channel so that the wrinkle generation process focuses on the line energy feature.
  • the line energy feature enhancement may include subtracting the average value of all features in the feature map on each channel to obtain high-frequency component features in these feature maps, and then the line energy features may be extracted from these high-frequency component features.
  • determining the line energy feature of each pixel in each channel feature map based on the multi-channel feature map may include:
  • the line energy feature of each pixel is obtained according to the line direction energy, and the line energy feature includes the maximum line direction energy of the pixel in the plurality of predetermined directions and a predetermined direction corresponding to the maximum line direction energy.
  • a linear convolutional layer may include several Gaussian Modified Finite RAdon Transform (MFRAT) kernels along different directions (i.e., the multiple predetermined directions mentioned above).
  • MFRAT Gaussian Modified Finite RAdon Transform
  • the line energy feature of the feature map can be obtained according to the maximum response operation. Specifically, for each pixel in the feature map, the maximum line direction energy of the pixel in the plurality of predetermined directions and a predetermined direction corresponding to the maximum line direction energy are selected as the line energy feature of the pixel.
  • the palmprint lines usually have certain directional characteristics, and the line direction energy can be used to describe the line direction information at different pixels in the palmprint image, for example, to quantify the line direction information at different pixels (or specific areas) and convert the line direction information into a set of numerical features to represent the line direction energy of the pixel.
  • a dilated convolution based on the Gaussian-MFRAT kernel can be used to reduce the computational and time costs.
  • the line energy feature of the feature map can be multiplied by the preset learning parameter S, and the product is added to the original feature map to obtain a feature map with enhanced line energy features.
  • the Gaussian-MFRAT kernel can be obtained by improving the traditional MFRAT method.
  • FIG4B is a schematic diagram showing a comparison between the Gaussian-MFRAT kernel 420 according to an embodiment of the present application and the traditional MFRAT filter 410.
  • the traditional MFRAT method uses a linear filter with a constant value, which is sensitive to noise or small changes, while the Gaussian-MFRAT kernel in this application is calculated as follows:
  • (x,y) ⁇ L( ⁇ ) represents the coordinates on the kernel
  • (x 0 ,y 0 ) represents the center point of the kernel
  • L( ⁇ ) represents a line with an angle ⁇ defined on the two-dimensional image plane (as shown by the white line on the black background in FIG. 4B )
  • ⁇ is a hyperparameter.
  • f(x, y) 0.
  • 12 Gaussian-MFRAT kernels may be designed, with a size of 31 ⁇ 31, ⁇ ranging from 0° to 165°, and an interval of 15°.
  • line direction energy can be extracted from the feature map along these different predetermined directions, and the line energy feature of each pixel in the feature map includes the maximum line direction energy selected from these line direction energies along different predetermined directions.
  • FIG. 4B also shows the filtering results 411 and 421 for the line energy feature generated based on the MFRAT filter and the Gaussian-MFRAT kernel.
  • the line energy feature extraction result based on the Gaussian-MFRAT kernel i.e., filtering result 421) is clearer and has less noise.
  • the Gaussian-MFRAT kernel 420 can be used to replace the traditional MFRAT filter 410, avoiding the use of the inefficient response suppression denoising strategy in the traditional MFRAT filter 410, simplifying the enhancement operation of the line energy feature, and achieving the differentiability of the enhancement operation.
  • ⁇ i represents the mean of Xi
  • f MAX represents the maximum response operation
  • N k the total number of Gaussian-MFRAT kernels
  • s i represents the preset learning parameter of the ith channel, which is used to adjust the feature enhancement degree of the ith channel.
  • a PCE image with realistic palm lines ie, wrinkles
  • the palm lines appear as different line types depending on the line energy characteristics of each pixel.
  • the line direction energy of each pixel in the palm line may include the line direction energy of the pixel in all directions.
  • the line direction energy of each pixel in all directions may be determined based on a feature-enhanced multi-channel feature map of a Bezier curve, and these line direction vectors may be represented by different line types, thereby achieving simulation of real wrinkles.
  • generating a palm line energy map with wrinkle information based on the Bezier curve may include: generating a palm line energy map based on the Bezier curve using a pre-trained first generator.
  • a first generator 304GB ⁇ P can be used to convert the Bezier curve domain (B) to the PCE domain (P).
  • the main structure of GB ⁇ P may include an encoder-decoder network based on a residual block (RB), and its specific structure and operation will be described below with reference to FIG6 .
  • a simulated palmprint image with detailed texture information may be generated based on the palmprint line energy map.
  • the simulated palmprint image has palmprint lines that are consistent with the palmprint line energy map and has detailed texture information.
  • preserving wrinkle content can be decoupled from generating detail textures.
  • generating a simulated palmprint image based on the palmprint line energy map may include: using multiple control vectors to generate the simulated palmprint image with multiple detail texture information based on the palmprint line energy map.
  • the palm lines in the palm line energy map are used as identity information to indicate the identity of the palm line image. That is, a realistic palm line image with detailed texture information can be generated while retaining the palm lines in the palm line energy map, wherein the palm lines retained in the palm line energy map can be interpreted as maintaining consistent identity information, because the palm lines can be used to indicate identity (ID).
  • detail texture information may be used to describe subtle texture features in a palm print.
  • the detail texture information may include one or more of light information, shadow information, and skin texture information.
  • skin texture information may include spots and fine textures on the skin, which are usually formed based on factors such as tiny details of the skin, arrangement and tissue structure of skin cells, and can be used to generate more realistic palm print images.
  • detail texture information may also include information related to the style of the palmprint image, such as light, shadow, and other information of the generated palmprint image, so as to achieve diversified palmprint image generation.
  • the generation of detail texture information may be achieved using a control vector.
  • the control vector may be a random noise vector.
  • a control vector 306 may participate in the generation of a simulated palm print image 308 from a PCE image 305, and optionally, the control vector may be a random noise vector (e.g., a Gaussian white noise conforming to a standard normal distribution) to control the generation of a diversified palm print image in the second generation stage.
  • a random noise vector e.g., a Gaussian white noise conforming to a standard normal distribution
  • control vectors may also be generated to generate diversified palm prints, and the present application does not impose any limitation on this.
  • generating a realistic palmprint image with detailed texture information based on the palmprint line energy map may include: generating a realistic palmprint image based on the palmprint line energy map using a pre-trained second generator.
  • a second generator 307GP ⁇ R can be used to convert the PCE image 305 of the PCE domain (P) into the palmprint image domain (R) to generate realistic palmprints using the PCE image 305 as an ID condition.
  • a random noise vector can be used as a control vector 306 and input into GP ⁇ R in the form of a latent vector to reproduce various detail textures such as light, shadow, and skin texture.
  • the first generator and the second generator may be jointly trained using real palmprint images. Therefore, next, the joint training process of the model used in the palmprint image generation method of the present application will be introduced with reference to FIGS. 5-8.
  • FIG5 is a schematic diagram illustrating domain conversion in joint training according to an embodiment of the present application.
  • the joint training process of the present application also involves a conversion from the Bessel palmprint domain to the PCE domain in the first generation stage, and a conversion from the PCE domain to the palmprint image domain in the second generation stage.
  • the PCE image sample 5031 is converted from the Bezier curve sample 501. Since there is no supervision information, the Convert to unpaired domain conversion 502.
  • a cycle conditional generation model is introduced.
  • the real palmprint image 506 is used as the supervision information.
  • the PCEE 505 is used to extract the real PCE image 5032 from the real palmprint image 506. Then the PCE image samples 5031 and 5032 are paired to generate a realistic palmprint image (not shown) in the palmprint image domain, thereby achieving the purpose of supervised learning using real palmprint images.
  • the first generator may take Bezier curve samples as input and palm line energy map samples as output, wherein the palm line energy map samples include wrinkle information.
  • FIG6 is a schematic diagram showing the first generation stage in joint training according to an embodiment of the present application.
  • the first generator GB ⁇ P to be trained can be used to generate the PCE image sample 602 in the PCE domain.
  • Figure 6(b) shows an exemplary structure of the first generator GB ⁇ P , wherein the main structure of the first generator GB ⁇ P may be an encoder-decoder network based on the residual block RB 610, and according to an embodiment of the present application, the first generator may include a line energy feature enhancement block (Line Feature Enhancement Block, LFEB) 620 for enhancing the line energy features of the multi-channel feature map of the Bezier curve sample 601.
  • Line Feature Enhancement Block Line Feature Enhancement Block
  • each encoder structure of the first generator GB ⁇ P may include a LFEB 620, a convolution layer (e.g., Conv 7x7, Conv 3x3, etc.) and an activation layer (e.g., BN+ReLU), and each decoder structure may include a deconvolution layer (e.g., DeConv 3x3) and an activation layer (e.g., BN+ReLU, Tanh).
  • the residual block RB 610 includes a convolution layer (e.g., Conv 3x3), an activation layer (e.g., BN+ReLU, BN), and a discard layer (e.g., Dropout).
  • the line energy feature enhancement block LFEB 620 can be used to implement the line energy feature enhancement processing described above with reference to step S202, as shown in the above formula (2), which may include:
  • the maximum line energy feature among the extracted multiple line energy features is used as the line energy feature of the feature map
  • the line energy feature enhancement block LFEB 620 is a lightweight plug-and-play block that can facilitate domain transmission and improve recognition performance.
  • a real PCE image 604 is obtained based on a real palmprint image 603 , so as to be used for supervising the training of the PCE image sample 602 .
  • a palmprint line energy extractor (PCE Extractor, PCEE) is used to extract the corresponding real PCE image from the real palmprint image.
  • a PCEE for extracting a PCE image from a real palmprint image can be designed to generate supervisory information for joint training, so that the generated palmprint image is more realistic.
  • FIG. 7 is a schematic diagram showing a palmprint line energy extractor PCEE according to an embodiment of the present application.
  • the palmprint line energy extractor PCEE of the present application may include a mean filter 710, a linear convolution layer 720, a maximum response operation layer 730, and an adaptive binarization layer 740, taking a real palmprint image 700 as input, and finally generating a binarized PCE image 750.
  • using a palm line energy extractor to extract a real palm line energy map from the real palm line image may include:
  • the real palm line energy map is obtained through binarization processing.
  • a mean filter 710 is applied to subtract the filtering result from the original real palmprint image 700 to obtain a high-frequency component;
  • a linear convolution layer 720 is used to obtain the line energy feature in the real palmprint image 700, that is, the line distribution of the real palmprint line, wherein the linear convolution layer may be composed of several Gaussian-MFRAT kernels along different directions.
  • the maximum line direction energy can be determined as the line energy feature of the pixel.
  • the final PCE image may be obtained using an adaptive binarization process 740.
  • a binarization threshold T may be set based on the top 10% of the values of the line energy features on the entire image.
  • the second generator in joint training, can take a real palmprint line energy map of a real palmprint image as input, and take a simulated palmprint image corresponding to the real palmprint image as output, wherein the simulated palmprint image includes detailed texture information.
  • FIG8 is a schematic diagram showing the second generation stage in joint training according to an embodiment of the present application.
  • the second generation stage in the joint training utilizes the real palmprint image 801 to participate in the training, wherein, similar to the above, PCEE is also used to extract the real PCE image 802 from the real palmprint image 801.
  • the real PCE image f PCEE (A) is generated from the real palmprint image A through PCEE, and then the real PCE image f PCEE (A) can be input into the second generator G P ⁇ R to be trained to obtain a simulated palmprint image A * 803 having palmprint lines consistent with the real PCE image f PCEE (A).
  • the encoder E can be used to map the input real palmprint image A to a vector with mean ⁇ Q and variance
  • A) can obey the normal distribution
  • a latent vector ie, control vector for controlling the generation of texture information is generated based on the latent space.
  • the divergence of the latent vector can be constrained during the training phase, for example, the distribution of the latent space can be made Close to the standard normal distribution N(0,1), that is, the latent space is approximated to the standard normal space, so that the random noise z ⁇ N(0,1) can be easily sampled as a latent vector to generate diverse palmprint details.
  • the generated simulated palmprint image 803 and the real palmprint image 801 can be expanded using a data augmentation module AUG before feeding these images to the discriminator D to generate more diverse training samples, thereby improving the generalization ability and robustness of the discriminator D. Even if only a small number of samples are used for training, the overfitting risk of the discriminator D can be reduced, allowing it to better adapt to different palmprint image samples.
  • the generation process of the first generator and the second generator can be supervised based on the real PCE image, wherein the palm line energy extractor can be jointly trained with the first generator and the second generator.
  • a real PCE image 604 can be used to supervise the generation of a PCE image sample 602.
  • a real PCE image 604 extracted from a real palmprint image 603 can be used as an adversarial sample to adopt an adversarial loss to drive the result generated by the first generator to be closer to the real PCE image 604, thereby optimizing the generation quality of the PCE image.
  • a loop structure with PCEE can be used in the second generation stage in the joint training to remap the generated simulated palmprint image A * back to the PCE domain fPCEE (A * ), that is, to obtain a simulated PCE image 804, and by minimizing the gap between the real PCE image 803 and the simulated PCE image 804, it is ensured that the simulated palmprint image has consistent identity information with the original real palmprint image.
  • the distortion of the generated simulated palmprint image A * can be constrained by minimizing the L1 distance between fPCEE (A) and the generated fPCEE (A * ), so as to strictly retain the ID information of the input fPCEE (A), and the discriminator D can be applied to strengthen the authenticity of the generated simulated palmprint image A * .
  • the loss function of the joint training may include a first loss function associated with the first generator and a second loss function associated with the second generator;
  • the first loss function includes: a contrast loss for maintaining the structural consistency between the palm line energy map sample and the Bezier curve sample, and an adversarial loss for making the palm line energy map sample similar to the real palm line energy map;
  • the second loss function includes: a distribution control loss for making the distribution of the control vector close to a standard normal distribution, an identity consistency loss between the identity of the simulated palmprint image and the identity of the real palmprint image, a distortion loss of the simulated palmprint image, and a loss for strengthening the authenticity of the simulated palmprint image.
  • the joint training of the present application may include joint training of the first generator, the second generator and the PCEE, wherein the PCEE serves as a connection between the first generator and the second generator. Therefore, the loss function of the entire training process may include a loss function corresponding to the generation process of the first generator and the generation process of the second generator.
  • the loss function corresponding to the generation process of the first generator may include a contrast loss L CL for maintaining structural consistency and an adversarial loss L adv for making the palmprint image more realistic.
  • a contrastive loss can be used. Specifically, a series of tiles can be cropped from both B and GB ⁇ P (B). For a query tile q in GB ⁇ P (B), the corresponding tile at the same position in B can be used as the positive sample k + of the query tile q, and other tiles in B can be used as the negative samples of the query tile Therefore, the contrastive loss function can be constructed in a way that the positive samples are closer together and the negative samples are farther apart.
  • the contrastive loss L CL applying the normalized mutual information neural estimation (InfoNCE) loss function can be expressed as follows:
  • ⁇ is the temperature hyperparameter
  • the loss function corresponding to the generation process of the first generator can be expressed as follows:
  • f PCEE represents the PCEE processing performed on image A
  • A represents a random real palmprint image, and are two weights.
  • the loss function corresponding to the generation process of the second generator may include a distribution control loss L KL for making the distribution of the control vector N(z) close to the standard normal distribution Q(z
  • a distribution control loss L KL for making the distribution of the control vector N(z) close to the standard normal distribution Q(z
  • an identity consistency loss L cyc for maintaining the identity of the simulated palmprint image 803 generated by the second generator based on the real PCE image 802 and the identity of the real palmprint image 801
  • a generation loss LG for constraining the distortion of the simulated palmprint image 803 generated by the second generator and enhancing the authenticity.
  • the distribution control loss L KL can be expressed as follows:
  • identity consistency loss Lcyc can be expressed as follows:
  • the generation loss LG may include a distortion loss for the simulated palm print image 803. and the loss of authenticity of the simulated palmprint image 803 for the discriminator Among them, T() represents the processing of the data enhancement module T. Therefore, the generation loss LG can be expressed as follows:
  • the loss function corresponding to the generation process of the second generator can be expressed as follows:
  • the same experimental data set and open set evaluation scheme as in the Bessel palmprint generation method and the RPG-Palm (Realistic Pseudo-data Generation-Palm) palmprint generation method can be followed.
  • the performance of the recognition model pre-trained based on the palmprint images generated by various palmprint generation methods can be evaluated according to TAR and FAR, where TAR and FAR represent "True Acceptance Rate” and "False Acceptance Rate", respectively. That is, TAR can represent the proportion of correctly identified samples that are correctly accepted, which can also be understood as the accuracy of the recognition model, while FAR represents the proportion of incorrectly identified samples that are incorrectly accepted, which can also be understood as the false recognition rate of the recognition model.
  • the FID (Fréchet Inception Distance) metric can be used to evaluate the quality of the generated palmprint images.
  • the Bessel palmprint generation method and the RPG palmprint generation method can be followed.
  • 13 public data sets are used, which can come from various devices, with a total of 3,268 IDs and 59,162 images.
  • the detect-then-crop scheme can be followed to extract the region of interest (ROI).
  • 4000 body images can be generated according to the Bessel palmprint generation method and the RPG palmprint generation method.
  • Each identity has 100 samples by default.
  • ⁇ can be set to 1.0, 1.0, and 1.0, and the learning rate is 0.0002 in the first 30 training epochs and decays linearly to 1e-6 in the last 30 training epochs.
  • the second generation phase and They are set to 1.0, 10.0, 0.01, and 1.0 respectively, and the learning rate is 0.0002 in the first 50 training cycles and linearly decays to 1e-8 in the last 50 training cycles.
  • the Adaptive Moment Estimation (Adam) optimizer parameters are set to (0.5, 0.99).
  • the resolution of all images in the above training can be set to 256 ⁇ 256.
  • the same recognition model skeleton as the palmprint recognition model corresponding to the Bessel palmprint generation method namely ResNet (residual network) 50 and MobileFaceNet (mobile face recognition network)
  • ResNet residual network
  • MobileFaceNet mobile face recognition network
  • SGD stochastic gradient descent
  • the method described in the embodiment of the present application can improve the RPG palmprint generation method with a clear margin, and achieves the highest performance level in the settings where the ratio of training ID to test ID is 1:1 and 1:3 respectively.
  • the improvement of the method described in the embodiment of the present application in the setting where the ratio of training ID to test ID is 1:3 is greater than that in the setting where the ratio is 1:1, that is, the method described in the embodiment of the present application has significant effectiveness in the case of less real data.
  • the model with different numbers of training identities can be tested under an open set protocol with a ratio of training ID to test ID of 1:1. Specifically, a total of 4000 pseudo IDs were synthesized in the verification process, and each ID contained 100 pseudo palmprints.
  • the same MobileFaceNet was used as the recognition model skeleton for different methods, and the quantitative results are shown in Table 2 below.
  • the palmprint generation performance can be compared using four generation methods, pix2pixHD, CycleGAN, BicycleGAN, and RPG-Palm, all of which are retrained using 40 real IDs and unpaired data from RPG-Palm.
  • the quantitative results are shown in Table 3.
  • the palmprint recognition model pre-trained based on the method described in the embodiment of the present application is superior to other methods.
  • the method described in the embodiment of the present application can achieve a FID score of 40.3, showing a significant improvement.
  • FIG9 is a comparison diagram showing palmprint generation results using different palmprint generation methods according to an embodiment of the present application.
  • (a) corresponds to the Bézier palm generation method
  • (b) corresponds to the pix2pixHD method
  • (c) corresponds to the CycleGAN method
  • (d) corresponds to the BicycleGAN method
  • (e) corresponds to the RPG-Palm method
  • (f) corresponds to the PCE image
  • (g)-(j) correspond to the diversified palmprint images generated by the method described in the embodiment of the present application.
  • the palm print image generated by RPG-Palm exhibits severe blurring and inconsistent lines, while the method described in the embodiment of the present application still maintains overall clarity and ID consistency.
  • the method described in the embodiment of the present application can restore better detail information about the palm print line, including changes in thickness and the interweaving of multiple wrinkles, rather than just migrating Bezier curves.
  • the results of the pix2pixHD, CycleGAN, and BicycleGAN methods show more serious blurring problems than RPG-Palm.
  • the palmprint image generation method of the present application can also apply the proposed LFEB to the palmprint recognition model, for example, incorporating a plug-and-play LFEB before the first convolutional layer of the skeleton of the palmprint recognition model to enhance the line energy characteristics of the input palmprint image.
  • the performance verification may also include a study of ablation.
  • the main components of the method described in the embodiment of the present application may include PCEE, a data amplification (DA) module for few-sample training, an improved Bezier curve synthesis, a generation model with LFEB, and a recognition model with LFEB, which can be represented by "P", "A”, “I”, “G+L” and “R+L” respectively.
  • DA data amplification
  • LFEB generation model with LFEB
  • recognition model with LFEB which can be represented by "P", "A”, “I”, “G+L” and "R+L” respectively.
  • 40 IDs can be used to train ablation experiments, and the test set is fixed under an open set protocol with a ratio of 1:1 between training ID and test ID. The results of the ablation experiment are shown in Table 4 below.
  • the improvement to the Bezier curve also achieves better performance by introducing a more reasonable palmprint line distribution.
  • Fig. 10 is a diagram showing the results of the effectiveness verification of the line energy feature enhancement block according to an embodiment of the present application.
  • Figure 10 visualizes the features of the middle layer 1 block of MobileFaceNet with and without LFEB, where Figure 10(a) is the input palmprint image, Figure 10(b) is the palmprint image after LFEB, Figure 10(c) corresponds to the feature visualization without LFEB, and Figure 10(d) corresponds to the feature visualization with LFEB. Therefore, it can be seen that the model focuses on both the palmprint lines and non-line areas without LFEB, and by adding the LFEB component, the model can be biased to focus on the palmprint lines.
  • a Bezier curve can be generated by using control points determined from a predetermined palmprint curve template, and the Bezier curve can be converted into a palmprint line energy map with wrinkle information, wherein the wrinkle information includes the same line distribution as the Bezier curve but a different line type, and its line type is determined by the line energy feature of each pixel, and then a palmprint image with texture information is further generated based on the palmprint line energy map with the wrinkle information, and the palmprint lines of the palmprint image are consistent with the wrinkle information of the palmprint line energy map, thereby generating a realistic palmprint image with realistic wrinkles and texture.
  • the palmprint line energy domain as the intermediate domain connecting the Bezier palmprint domain and the palmprint image domain
  • the direct generation of palmprint images with wrinkle information and texture information from the Bezier curve is avoided, thereby reducing the difficulty of generating palmprint images.
  • detailed texture information is generated while ensuring that the palmprint image has consistent palmprint lines, so that realistic palmprint images with diverse textures can be generated while retaining the same identity information. Therefore, the dependence on real data is reduced, which is suitable for palmprint recognition training in the absence of large-scale palmprint datasets.
  • FIG. 11 is a schematic diagram showing a palmprint image generating device 1100 according to an embodiment of the present application.
  • the palmprint image generating device 1100 may include a curve generating module 1101 , a wrinkle generating module 1102 , and a texture generating module 1103 .
  • the curve generation module 1101 may be configured to determine a plurality of control points based on a predetermined palmprint curve template, and generate a Bezier curve based on the plurality of control points.
  • the curve generation module 1101 may perform the operations described above with reference to step S201.
  • the wrinkle generation module 1102 may be configured to generate a palm line energy map with wrinkle information based on the Bezier curve, wherein the palm line energy map includes palm lines having the same line distribution as the Bezier curve but different line types, wherein the line types are used to describe the wrinkle information and correspond to the line direction energy of each pixel on the palm line.
  • the wrinkle generation module 1102 may perform the operations described above with reference to step S202.
  • the texture generation module 1103 may be configured to generate a simulated palm print image with detailed texture information based on the palm print line energy map.
  • the texture generation module 1103 may perform the operation described above with reference to step S203.
  • the wrinkle generation module 1102 is configured to perform feature extraction on the Bezier curve to obtain a multi-channel feature map of the Bezier curve; determine the line energy features of each pixel in each channel feature map based on the multi-channel feature map; enhance the line energy features of each pixel to generate an enhanced multi-channel feature map; and generate the palm print line energy map based on the enhanced multi-channel feature map.
  • the wrinkle generation module 1102 is configured to perform the following processing for each channel feature map:
  • the linear convolution layer includes a Gaussian modified finite Radon transform kernel along a plurality of predetermined directions, and the line direction energy includes the line direction energy of the pixel along the plurality of predetermined directions;
  • the line energy feature of each pixel is obtained according to the line direction energy, and the line energy feature includes the maximum line direction energy of the pixel in the plurality of predetermined directions and a predetermined direction corresponding to the maximum line direction energy.
  • the texture generation module 1103 is configured to generate the simulated palm print image with multiple detail texture information based on the palm print line energy map using multiple control vectors.
  • the palm lines in the palm line energy map are used as identity information to indicate the identity of the simulated palm print image.
  • control vector is a random noise vector
  • detail texture information includes one or more of light information, shadow information and skin texture information.
  • the wrinkle generation module 1102 is configured to generate the palm line energy map based on the Bezier curve using a pre-trained first generator
  • the texture generation module 1103 is configured to generate the simulated palm print image based on the palm print line energy map using a pre-trained second generator;
  • the first generator and the second generator are jointly trained using real palmprint images.
  • the first generator takes a Bezier curve sample as input and a palm line energy map sample as output, wherein the palm line energy map sample includes wrinkle information;
  • the second generator takes the real palmprint line energy map of the real palmprint image as input, and takes the simulated palmprint image corresponding to the real palmprint image as output, wherein the simulated palmprint image includes detail texture information;
  • the palm line energy extractor is jointly trained with the first generator and the second generator.
  • the extracting the real palmprint line energy map from the real palmprint image using a palmprint line energy extractor comprises:
  • the real palm line energy map is obtained through binarization processing.
  • the first generator comprises a line energy feature enhancement block for enhancing the line energy features of the multi-channel feature map of the Bezier curve sample.
  • the loss function of the joint training includes a first loss function associated with the first generator and a second loss function associated with the second generator;
  • the first loss function includes: a contrast loss for maintaining the structural consistency between the palm line energy map sample and the Bezier curve sample, and an adversarial loss for making the palm line energy map sample similar to the real palm line energy map;
  • the second loss function includes: a distribution control loss for making the distribution of the control vector close to a standard normal distribution, an identity consistency loss between the identity of the simulated palmprint image and the identity of the real palmprint image, a distortion loss of the simulated palmprint image, and a loss for strengthening the authenticity of the simulated palmprint image.
  • the palm print curve template is predetermined based on statistical information obtained from real human palm print lines;
  • the curve generating module 1101 is configured to determine the region range for generating the plurality of control points based on the palmprint curve template; and perform sampling within the region range to obtain the plurality of control points.
  • Fig. 12 shows a schematic diagram of an electronic device 2000 according to an embodiment of the present application.
  • the electronic device 2000 may include one or more processors 2010 and one or more memories 2020.
  • the memory 2020 stores computer readable codes, and when the computer readable codes are run by the one or more processors 2010, the palmprint image generation method described above may be executed.
  • the processor in the embodiment of the present application can be an integrated circuit chip with signal processing capabilities.
  • the above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
  • DSP digital signal processor
  • ASIC application-specific integrated circuit
  • FPGA field-programmable gate array
  • the disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed.
  • the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can be an X86 architecture or an ARM architecture.
  • various example embodiments of the present application can be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects can be implemented in hardware, while other aspects can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device.
  • firmware or software that can be executed by a controller, microprocessor, or other computing device.
  • the method or apparatus according to the embodiment of the present application can also be implemented with the aid of the architecture of the computing device 3000 shown in FIG13.
  • the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input/output component 3060, a hard disk 3070, and the like.
  • the storage device in the computing device 3000 such as the ROM 3030 or the hard disk 3070, can store various data or files used for processing and/or communication of the palmprint image generation method provided in the present application, as well as program instructions executed by the CPU.
  • the computing device 3000 may also include a user interface 3080.
  • the architecture shown in FIG13 is only exemplary. When implementing different devices, one or more components in the computing device shown in FIG13 may be omitted according to actual needs.
  • a computer-readable storage medium is also provided.
  • Computer-readable instructions are stored on the computer storage medium.
  • the palm print image generation method according to the embodiment of the present application described with reference to the above figures can be executed.
  • the computer-readable storage medium in the embodiment of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories.
  • the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM) or a flash memory.
  • the volatile memory can be a random access memory (RAM), which is used as an external cache.
  • RAM random access memory
  • DRAM dynamic random access memory
  • SDRAM synchronous dynamic random access memory
  • DDRSDRAM double data rate synchronous dynamic random access memory
  • ESDRAM enhanced synchronous dynamic random access memory
  • SLDRAM synchronous connection dynamic random access memory
  • DR RAM direct memory bus random access memory
  • the embodiment of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the palmprint image generation method according to the embodiment of the present application.
  • Embodiments of the present application provide a palmprint image generation method, apparatus, device, and computer-readable storage medium.
  • the method provided in the embodiment of the present application can decouple the generation of palmprint wrinkles and textures by introducing the palmprint line energy domain, so as to convert the Bezier curve in the Bezier palmprint domain into the palmprint line energy domain.
  • the realistic wrinkles are generated by converting the palmprint line energy map in the palmprint line energy domain into the palmprint image in the palmprint image domain to generate realistic texture, thereby reducing the difficulty of generating realistic palmprint images from Bezier curves and realizing diversified realistic palmprint image generation.
  • the method provided in the embodiment of the present application generates a Bezier curve using control points determined from a predetermined palmprint curve template, and converts the Bezier curve into a palmprint line energy map with wrinkle information, wherein the wrinkle information includes the same line distribution as the Bezier curve but a different line type, and the line type is determined by the line energy feature of each pixel, and then further generates a palmprint image with texture information based on the palmprint line energy map with the wrinkle information, and the palmprint lines of the palmprint image are consistent with the wrinkle information of the palmprint line energy map, thereby generating a simulated palmprint image with realistic wrinkles and texture.
  • the method provided in the embodiment of the present application introduces a palmprint line energy domain as an intermediate domain connecting the Bezier palmprint domain and the palmprint image domain, thereby avoiding direct generation of a palmprint image with wrinkle information and texture information from a Bezier curve, thereby reducing the difficulty of generating a palmprint image.
  • a palmprint line energy domain as an intermediate domain connecting the Bezier palmprint domain and the palmprint image domain, thereby avoiding direct generation of a palmprint image with wrinkle information and texture information from a Bezier curve, thereby reducing the difficulty of generating a palmprint image.
  • detailed texture information is generated while ensuring that the palmprint image has consistent palmprint lines, so that a realistic palmprint image with diversified textures can be generated while retaining the same identity information, thereby reducing dependence on real data and being suitable for palmprint recognition training in the absence of a large-scale palmprint data set.
  • each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains at least one executable instruction for realizing the specified logical function.
  • the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.
  • each box in the block diagram and/or flowchart, and the combination of the boxes in the block diagram and/or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
  • various example embodiments of the present application can be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects can be implemented in hardware, while other aspects can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device.
  • firmware or software that can be executed by a controller, microprocessor, or other computing device.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Human Computer Interaction (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Image Processing (AREA)

Abstract

本申请实施例提供了一种掌纹图像生成方法、装置、设备、存储介质及程序产品。该方法包括：基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线；基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应；及，基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。

Description

掌纹图像生成方法、装置、设备、存储介质及程序产品
本申请要求于2023年11月17日提交中国专利局、申请号为202311546177.7、申请名称为“掌纹图像生成方法、装置、设备和存储介质”的中国专利申请的优先权。
技术领域
本申请涉及人工智能领域，更具体地，涉及一种掌纹图像生成方法、装置、设备、存储介质及程序产品。
发明背景
在掌纹识别领域中，可用于训练和评估的大规模掌纹数据集非常有限，这种情况严重阻碍了掌纹识别技术的发展和性能提升。掌纹是指人手掌内部的皮肤纹路，具有独特性和稳定性，可以用于个体识别和身份验证等。然而，由于掌纹数据的获取和标注相对复杂和耗时，导致可用的掌纹数据集数量有限，特别是大规模的数据集。
在深度学习等领域，数据量对于模型的训练和性能至关重要。大规模数据集可以提供更多的样本和变化，帮助模型更好地学习和泛化。然而，由于大规模掌纹数据集的缺乏，传统的深度学习方法在掌纹识别中的表现受到限制。
尽管大规模掌纹数据集的稀缺性是一个挑战，但在收集大规模掌纹数据集时必须注意保护用户隐私。掌纹是个人身体特征的一部分，具有敏感性和隐私性。因此，在收集和使用掌纹数据时，必须遵守相关的隐私法律和规定，并采取适当的安全措施来保护用户的隐私。
因此，需要一种高效的掌纹图像生成方法，使得能够生成多样化且逼真的掌纹图像，同时保护用户隐私。
发明内容
为了解决上述问题，本申请通过引入新的掌纹线能量(Palm Crease Energy，PCE)域，首先将贝塞尔曲线转换至掌纹线能量域以生成具有逼真皱褶(crease)的掌纹线能量图，继而基于掌纹线能量图生成具有逼真纹理(texture)的拟真掌纹图像，从而生成多样化且逼真的掌纹图像。
本申请实施例提供了一种掌纹图像生成方法、装置、设备、存储介质及程序产品。
一方面，本申请实施例提供了一种掌纹图像生成方法，由电子设备执行，包括：
基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线；
基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应；及，
基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。
另一方面，本申请实施例提供了一种掌纹图像生成装置，包括：
曲线生成模块，被配置为基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线；
皱褶生成模块，被配置为基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应；及，
纹理生成模块，被配置为基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。
另一方面，本申请实施例提供了一种电子设备，包括：一个或多个处理器；以及一个或多个存储器，其中，所述一个或多个存储器中存储有计算机可执行程序，当由所述处理器执行所述计算机可执行程序时，执行如上所述的掌纹图像生成方法。
另一方面，本申请实施例提供了一种计算机可读存储介质，其上存储有计算机可执行指令，所述指令在被处理器执行时用于实现如上所述的掌纹图像生成方法。
另一方面，本申请实施例提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行根据本申请实施例的掌纹图像生成方法。
附图简要说明
图1是示出贝塞尔曲线与真实掌纹在皱褶分布和纹理方面的显著差距的示意图；
图2是示出根据本申请实施例的掌纹图像生成方法的流程图；
图3是示出根据本申请实施例的掌纹图像生成系统的示意图；
图4A是示出根据本申请实施例的示例掌纹曲线模板的示意图；
图4B是示出根据本申请实施例的高斯-MFRAT核与传统的MFRAT滤波器的对比的示意图；
图5是示出根据本申请实施例的联合训练中的领域转换的示意图；
图6是示出根据本申请实施例的联合训练中的第一生成阶段的示意图；
图7是示出根据本申请实施例的掌纹线能量提取器的示意图；
图8是示出根据本申请实施例的联合训练中的第二生成阶段的示意图；
图9是示出根据本申请实施例的利用不同掌纹生成方法的掌纹生成结果的对比图；
图10是示出根据本申请实施例的针对线能量特征增强块的有效性验证的结果图；
图11是示出根据本申请实施例的掌纹图像生成装置的示意图；
图12示出了根据本申请实施例的电子设备的示意图；以及
图13示出了根据本申请实施例的示例性计算设备的架构的示意图。
实施方式
为了使得本申请目的、技术方案和优点更为明显，下面将参考附图详细描述根据本申请示例实施例。显然，所描述的实施例仅仅是本申请一部分实施例，而不是本申请全部实施例，应理解，本申请不受这里描述的示例实施例的限制。
在本说明书和附图中，具有基本上相同或相似步骤和元素用相同或相似的附图标记来表示，且对这些步骤和元素的重复描述将被省略。同时，在本申请描述中，术语“第一”、“第二”等仅用于区分描述，而不能理解为指示或暗示相对重要性或排序。
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请技术领域的技术人员通常理解的含义相同。本文中所使用的术语只是为了描述本发明实施例的目的，不是旨在限制本发明。
为便于描述本申请，以下介绍与本申请有关的概念。
本申请掌纹图像生成方法可以基于人工智能(Artificial intelligence，AI)来实现。人工智能是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，使机器具有感知、推理与决策的功能。具体地，在本申请实施例中，人工智能通过研究各种智能机器的设计原理与实现方法，使本申请掌纹图像生成方法实现如下功能：基于从掌纹曲线模板确定的控制点，生成具有逼真的皱褶(crease)和纹理(texture)的拟真掌纹图像。
本申请掌纹图像生成方法还可以基于计算机视觉(Computer Vision，CV)技术来实现。计算机视觉技术，能够从图像或者多维数据中获取信息。具体地，本申请掌纹图像生成方法可以利用CV技术，从贝塞尔曲线生成具有接近于真实掌纹的风格的掌纹线能量图，继而基于掌纹线能量图，生成具有多样化纹理信息的掌纹图像，以实现多样化的掌纹图像输出，用于诸如掌纹识别模型的预训练过程等。
本申请掌纹图像生成方法可以基于贝塞尔(Bezier)曲线来实现。贝塞尔曲线用于描述平滑的曲线，在本申请实施例中，可以将贝塞尔曲线应用于掌纹线的绘制。贝塞尔曲线的特点是可以通过控制点来控制曲线的形状，因此可以通过控制点来调整掌纹线的弯曲程度和曲线的形状。具体地，可以将手掌分成若干段，每一段用一条贝塞尔曲线来描述，通过调整控制点的位置和数量，可以得到不同形状的掌纹线。
综上所述，本申请实施例提供的方案涉及人工智能、计算机视觉等技术，下面将结合附图对本申请实施例进行进一步地描述。
图1是示出贝塞尔曲线与真实掌纹在皱褶分布和纹理方面存在显著差距的示意图。
掌纹作为一种稳定且隐私友好的生物特征识别技术，最近在识别应用中显示出巨大的潜力。近年来，基于深度学习的掌纹识别方法已成为主流的掌纹识别技术。基于深度学习的掌纹识别方法训练神经网络以提取具有改进的分类或配对损失的掌纹的特征。然而，基于深度学习的掌纹识别的研究和应用中的一个主要难点在于大规模掌纹数据集的稀缺，收集大规模掌纹数据集可能引发违反用户隐私的风险。为了克服这一问题，研究人员目前可以通过使用一些数据合成技术来生成模拟的掌纹数据，从而扩充数据集。
目前，一些掌纹生成方法已经被应用于在掌纹识别领域中产生伪掌纹样本。例如，贝塞尔掌纹(Bézier palm)生成方法使用参数化的贝塞尔曲线来合成伪掌纹线。然而，如图1所示，通过贝塞尔掌纹生成方法生成的贝塞尔掌纹101，在皱褶和纹理方面，与真实的掌纹图像102均存在显著差距，其无法反映真实掌纹的皱褶分布，也无法呈现真实掌纹中的各种细节纹理。因此，贝塞尔掌纹仍然需要一定数量的真实的掌纹数据用于微调。
此外，在样本数据有限的情况下，基于生成对抗网络(GAN)的模型，通常面临着鉴别器过拟合以及离散数据空间和连续隐分布之间的不平衡的挑战，这些问题导致保真度降低以及训练过程不稳定。
并且，一些使用少样本的方法，诸如数据增强、正则化和迁移学习等，在生成掌纹时，缺乏对身份的可控。
基于此，本申请实施例提供了一种使用连接贝塞尔掌纹域和真实掌纹图像域的中间域的方法，其中，引入了一种新的掌纹线能量(Palm Crease Energy,PCE)域作为该中间域。首先，将贝塞尔曲线转换至该掌纹线能量域，以生成具有逼真皱褶的掌纹线能量图(简称为PCE图像)，继而基于掌纹线能量图，生成具有逼真纹理的拟真掌纹图像，从而得到多样化且逼真的掌纹图像。
如图1所示，PCE图像103作为接近真实掌纹的中间状态，其与贝塞尔曲线101具有皱褶一致性，即其皱褶(即掌纹线)的线条分布与贝塞尔曲线101一致，并且与真实的掌纹图像102具有外观相似性，即该PCE图像103的掌纹线的外观与真实的掌纹线的外观一致。
本申请实施例所提供的方法，相比于传统的掌纹生成方法而言，将贝塞尔-真实(Bezier-Real)差分解为皱褶差和纹理差，从而降低了生成难度。具体而言，通过引入掌纹线能量域来对掌纹的皱褶和纹理的生成进行解耦，以通过将贝塞尔掌纹域的贝塞尔曲线转换至掌纹线能量域的掌纹线能量图来生成逼真的皱褶，并通过将掌纹线能量域的掌纹线能量图转换至掌纹图像域的掌纹图像来生成逼真的纹理，从而降低了从贝塞尔曲线生成拟真掌纹图像的难度，实现了多样化的拟真掌纹图像生成。
本申请实施例所提供的方法，利用从预先确定的掌纹曲线模板中确定的控制点来生成贝塞尔曲线，并将该贝塞尔曲线转换为具有皱褶信息的掌纹线能量图，其中该皱褶信息包括与贝塞尔曲线相同的线条分布但不同的线型，其线型由每个像素的线能量特征确定，继而进一步基于具有该皱褶信息的掌纹线能量图生成具有纹理信息的掌纹图像，该掌纹图像的掌纹线与掌纹线能量图的皱褶信息一致，从而生成具有逼真皱褶和纹理的拟真掌纹图像。
本申请实施例所提供的方法，通过引入掌纹线能量域作为连接贝塞尔掌纹域和掌纹图像域的中间域，避免从贝塞尔曲线直接生成具有皱褶信息和纹理信息的掌纹图像，降低了掌纹图像的生成难度，并且在从掌纹线能量图生成掌纹图像的过程中，在保证掌纹图像具有一致掌纹线的同时生成细节纹理信息，使得能够在保留相同的身份信息的情况下生成具有多样化纹理的拟真掌纹图像，因此降低了对真实数据的依赖，适用于缺乏大规模掌纹数据集的掌纹识别训练。
图2是示出根据本申请实施例的掌纹图像生成方法200的流程图。图3是示出根据本申请实施例的掌纹图像生成系统的示意图。
如图3所示，本申请实施例中，掌纹图像生成系统可以包括与第一生成器304相关的第一生成阶段310、以及与第二生成器307相关的第二生成阶段320。其中，该第一生成阶段310和该第二生成阶段320以PCE域的PCE图像305作为连接。
在第一生成阶段310中，贝塞尔掌纹生成器302基于控制点301生成贝塞尔曲线303，第一生成器304基于贝塞尔曲线303生成PCE图像305。
在第二生成阶段320中，第二生成器307基于PCE图像305以及控制向量306，生成拟真掌纹图像308。下面将参考图2-图4A对该生成过程进行详细介绍。
首先，如图2所示，在步骤S201中，可以基于预先确定的掌纹曲线模板，确定多个控制点，并基于多个控制点生成贝塞尔曲线。
可选地，由于贝塞尔曲线是基于多个控制点生成的曲线，本申请实施例中，掌纹图像生成方法可以利用来自人类皮肤纹理得到的先验知识来改进控制点的生成机制，例如，改进用于生成控制点的掌纹曲线模板。
根据本申请实施例，所述掌纹曲线模板可以是基于人类真实掌纹线得到的统计信息预先确定的。由于同时考虑了人类真实掌纹的多样性以及个体差异，所确定的掌纹曲线模板更具代表性。因此，基于该掌纹曲线模板确定的控制点的生成范围会更准确，从而可以生成更接近真实掌纹皱褶分布的贝塞尔曲线。
图4A是示出根据本申请实施例的示例掌纹曲线模板的示意图。如图4A所示，给出了基于人类真实掌纹线得到的统计信息划分的五种示例掌纹曲线模板，分别对应于最具代表性的五种人类掌纹线分布。在这五种示例掌纹曲线模板中，每种示例掌纹曲线模板都给出了用于控制点的生成范围，如虚线框所示。不同示例掌纹曲线模板可以具有不同数量的掌纹线，因此可以用不同数量的生成范围和控制点来确定。
应当理解，图4A中的示例掌纹曲线模板并不作为对本申请用于生成控制点的掌纹曲线模板的限制，本申请同样可以采用其他各种掌纹曲线模板。
根据本申请实施例，基于预先确定的掌纹曲线模板，确定多个控制点可以包括：
基于所述掌纹曲线模板，确定生成所述多个控制点的区域范围；
在所述区域范围内进行采样，得到所述多个控制点。
作为示例，采样的方法可以包括诸如随机采样等方法，以在保持掌纹的形状和风格的基础上，增加一定的随机性和个体差异，使生成的掌纹图像更加真实和多样化。也可以使用其他采样方法以用于实现不同的数据生成效果，本申请对此不作限制。
因此，通过基于人类真实掌纹线得到的统计信息来生成掌纹曲线模板，这些掌纹曲线模板可以提供控制点的更准确的生成范围，使得基于这些控制点生成的贝塞尔曲线更接近真实皱褶的分布，从而减少所生成的掌纹与真实掌纹之间的差异。
接下来，在步骤S202中，可以基于贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，掌纹线能量图包括与贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，线型用于描述皱褶信息、并与掌纹线上各个像素的线方向能量相对应。
可选地，如图3所示，在与步骤S202相对应的第一生成阶段中，可以将贝塞尔曲线303转换到PCE域，以生成PCE图像(即掌纹线能量图)305，贝塞尔曲线303和PCE图像305都是基于线条的二值图像。第一生成阶段的目标在于，将贝塞尔掌纹图像中的曲线线条转换为具有更接近于人类真实掌纹的皱褶(即掌纹线)的PCE图像。
其中，皱褶(即掌纹线)可以指掌纹中的深浅凹凸线条，通常是由于皮肤的折叠和弯曲而形成的，其用于描述掌纹的整体形状，例如手掌的主要纹路、弯曲的边缘等，并且其在掌纹中起到了分割和定义不同区域的作用，掌纹线的形状和分布可以用于诸如个体身份识别等领域。
因此，与贝塞尔曲线相比，PCE图像可以包括线条分布相同(即，掌纹的主要纹路的位置和走向一致)的掌纹线，但是PCE图像中的掌纹线具有不同的线型(例如，不同粗细和深浅的线条)，以用于模拟人类真实掌纹的皱褶。
可选地，掌纹线上使用的不同线型可以取决于掌纹线上相应位置的像素的线能量特征，其中，每个像素的线能量特征可以是通过对贝塞尔曲线进行线能量特征增强得到的，其可以描述该像素处线能量特征的方向分布情况。
具体地，根据本申请实施例，基于所述贝塞尔曲线，生成掌纹线能量图可以包括：
对所述贝塞尔曲线进行特征提取，获得所述贝塞尔曲线的多通道特征图；
基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征；
对各个像素的所述线能量特征进行增强，生成增强后的多通道特征图；及
基于所述增强后的多通道特征图，生成所述掌纹线能量图。
作为示例，贝塞尔曲线具备N个通道，通过将贝塞尔曲线处理成多维矩阵，可以获得每个通道的 特征图，例如，第i个通道的特征图可以表示为三维矩阵Xi∈Rh×w×c,i＝1,…,N。然后，可以对每个通道上的特征图进行线能量特征增强，以使皱褶生成的过程专注于线能量特征。
可选地，该线能量特征增强可以包括对每个通道上的特征图，减去该特征图中所有特征的平均值，以获得这些特征图中的高频分量特征，然后，可以从这些高频分量特征中提取线能量特征。
根据本申请实施例，基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征，可以包括：
对于每个通道特征图，执行以下处理：
利用线状卷积层，从该通道特征图中提取每个像素的线方向能量(line orientation energy)，其中，所述线状卷积层包括沿着多个预定方向的高斯改进有限拉东变换核，所述线方向能量包括该像素沿着所述多个预定方向上的线方向能量；
根据所述线方向能量，获得每个像素的所述线能量特征，所述线能量特征包括该像素在所述多个预定方向上的最大线方向能量以及与所述最大线方向能量相对应的预定方向。
例如，线状卷积层可以包括沿着不同方向(即上述多个预定方向)的若干高斯改进有限拉东变换(Modified Finite RAdon Transform，MFRAT)核。
然后，可以根据最大响应操作来获得该特征图的线能量特征。具体地，对于特征图中的每个像素，选择该像素在所述多个预定方向上的最大线方向能量以及与所述最大线方向能量相对应的一个预定方向作为该像素的线能量特征。
在掌纹图像中，掌纹线通常具有一定的方向特征，而线方向能量可以用于描述掌纹图像中不同像素处的线条方向信息，例如，量化不同像素处(或特定区域)的线条方向信息，将线条方向信息转换为一组数值特征，用于表示该像素的线方向能量。
此外，考虑到高斯-MFRAT核的大小通常很大，在本申请实施例中，可以采用基于高斯-MFRAT核的扩张卷积来减少计算和时间成本。
最后，可以将该特征图的线能量特征与预设的学习参数S相乘，并将乘积添加到原始特征图，从而得到线能量特征增强后的特征图。
可选地，上述高斯-MFRAT核可以是基于传统的MFRAT方法进行改进得到的。例如，具体地，图4B是示出根据本申请实施例的高斯-MFRAT核420与传统的MFRAT滤波器410的对比的示意图。
传统的MFRAT方法使用具有恒定值的线状滤波器，其对噪声或微小变化敏感，而本申请中的高斯-MFRAT核计算如下：
其中，(x,y)∈L(θ)表示核上的坐标，(x0,y0)表示核的中心点，L(θ)表示在二维图像平面上定义的角度为θ的线(如图4B中的黑色背景上的白色线条所示)，并且σ是超参数。另外，当时，f(x,y)＝0。例如，在本申请实施例中，可以设计12个高斯-MFRAT核，其大小为31×31，θ为0°到165°，间隔为15°。
如图4B所示，通过采用沿不同预定方向(例如，图4B中左下白线所示的6个不同预定方向)的高斯-MFRAT核420，可以从特征图中沿这些不同预定方向提取线方向能量，该特征图中的每个像素的线能量特征包括从这些沿不同预定方向的线方向能量中选择的最大的线方向能量。
图4B还示出了基于MFRAT滤波器和高斯-MFRAT核所生成的针对线能量特征的滤波结果411和421。其中，与MFRAT滤波器的滤波结果411相比，基于高斯-MFRAT核的线能量特征提取结果(即滤波结果421)更清晰并且具有更少的噪声。
因此，可以使用高斯-MFRAT核420来代替传统MFRAT滤波器410，避免使用传统MFRAT滤波器410中低效的响应抑制去噪策略，并且简化了线能量特征的增强操作，还实现了该增强操作的可微分性。
因此，作为示例，对于多通道特征图X中的第i个通道上的特征图Xi，对其线能量特征进行增强，可以计算如下：
其中，μi表示Xi的均值，fMAX表示最大响应操作，表示第k个高斯-MFRAT核，高斯-MFRAT核的总数为Nk，例如设置为12，并且si表示第i个通道的预设的学习参数，其用于调整第i个通道的特征增强程度。
因此，在第一生成阶段中，可以基于贝塞尔曲线生成具有逼真掌纹线(即皱褶)的PCE图像，其中，该掌纹线表现为不同的线型，其取决于每个像素的线能量特征。
根据本申请实施例，所述掌纹线中的每个像素的线方向能量可以包括该像素在所有方向上的线方向能量。可选地，每个像素在所有方向上的线方向能量可以基于贝塞尔曲线的经特征增强的多通道特征图确定，并且这些线方向向量可以利用不同线型来表示，从而实现对真实皱褶的模拟。
根据本申请实施例，基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图可以包括：基于所述贝塞尔曲线，利用预先训练的第一生成器，生成掌纹线能量图。
可选地，如图3所示，可以使用第一生成器304GB→P，将贝塞尔曲线域(B)转换到PCE域(P)，GB→P的主要结构可以包括基于残差块(RB)的编码器-解码器网络，其具体结构和操作将在下文参考图6进行介绍。
在生成PCE域中具有皱褶信息的PCE图像后，在步骤S203中，可以基于掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。
本步骤中，所述拟真掌纹图像具有与所述掌纹线能量图一致的掌纹线，并且具有细节纹理信息。
可选地，在第二生成阶段，可以将保留皱褶内容与生成细节纹理进行解耦。
根据本申请实施例，基于所述掌纹线能量图，生成拟真掌纹图像可以包括：使用多个控制向量，基于所述掌纹线能量图，生成具有多个细节纹理信息的所述拟真掌纹图像。
其中，所述掌纹线能量图中的掌纹线，作为身份信息，用于指示所述掌纹图像的身份。也就是说，可以在保留掌纹线能量图中的掌纹线的情况下，生成具有细节纹理信息的拟真掌纹图像，其中，保留掌纹线能量图中的掌纹线可以被解释为保持一致的身份信息，因为掌纹线可以用于指示身份(ID)。
根据本申请实施例，细节纹理信息可以用于描述掌纹中的细微纹理特征，例如，所述细节纹理信息可以包括光线信息、阴影信息以及皮肤纹理信息中的一种或多种。
其中，皮肤纹理信息可以包括皮肤上的斑点、细小的纹理等，其通常是基于皮肤的微小细节、皮肤细胞的排列和组织结构等因素形成的，可以用于生成更逼真的掌纹图像。
此外，除了皮肤纹理信息，细节纹理信息还可以包括与掌纹图像的风格相关的信息，诸如生成的掌纹图像的光线、阴影等信息，以实现多样化的掌纹图像生成。
可选地，细节纹理信息的生成可以利用控制向量来实现。
根据本申请实施例，所述控制向量可以为随机噪声向量。如图3所示，控制向量306可以参与从PCE图像305到拟真掌纹图像308的生成，并且可选地，该控制向量可以是随机噪声(例如，符合标准正态分布的高斯白噪声)向量，以在第二生成阶段控制生成多样化的掌纹图像。
应当理解，除了上述随机噪声外，还可以生成其他形式的控制向量，以用于生成多样化掌纹，本申请对此不作限制。
根据本申请实施例，基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像可以包括：基于所述掌纹线能量图，利用预先训练的第二生成器，生成拟真掌纹图像。
可选地，如图3所示，可以使用第二生成器307GP→R，将PCE域(P)的PCE图像305转换到掌纹图像域(R)，以使用PCE图像305作为ID条件来产生逼真的掌纹。其中，为了生成多样化的掌纹图像，可以将随机噪声向量作为控制向量306，以隐向量的方式，输入到GP→R中，以再现诸如光线、阴影和皮肤纹理等的各种细节纹理。
根据本申请实施例，所述第一生成器和所述第二生成器可以是利用真实掌纹图像联合训练的。因此，接下来，将参考图5-图8对本申请掌纹图像生成方法中使用的模型的联合训练过程进行介绍。
图5是示出根据本申请实施例的联合训练中的领域转换的示意图。
如图5所示，与以上参考图2-图4B描述的掌纹图像生成过程类似地，在本申请联合训练过程中，同样涉及第一生成阶段从贝塞尔掌纹域到PCE域的转换，以及第二生成阶段从PCE域到掌纹图像域的转换。
在第一生成阶段中，从贝塞尔曲线样本501转换得到PCE图像样本5031。由于没有监督信息，该 转换为不配对域转换502。
在第二生成阶段中，引入了一个循环条件生成模型(cycle conditional generation model)，以真实掌纹图像506作为监督信息，使用PCEE 505，从真实掌纹图像506中提取出真实PCE图像5032，然后将PCE图像样本5031与5032进行配对，生成掌纹图像域的拟真掌纹图像(未示出)，从而实现了使用真实掌纹图像进行监督学习的目的。
根据本申请实施例，在所述联合训练中，所述第一生成器可以以贝塞尔曲线样本为输入，以掌纹线能量图样本为输出，其中，掌纹线能量图样本中包括皱褶信息。
图6是示出根据本申请实施例的联合训练中的第一生成阶段的示意图。
如图6(a)所示，在联合训练中的第一生成阶段中，与上述关于第一生成阶段的描述类似地，可以基于贝塞尔掌纹域中的贝塞尔曲线样本601，利用待训练的第一生成器GB→P生成PCE域中的PCE图像样本602。
图6(b)给出了该第一生成器GB→P的示例性结构，其中，第一生成器GB→P的主要结构可以是基于残差块RB 610的编码器-解码器网络，并且根据本申请实施例，所述第一生成器可以包括线能量特征增强块(Line Feature Enhancement Block,LFEB)620，用于对贝塞尔曲线样本601的多通道特征图的线能量特征进行增强。
具体地，如图6(b)所示，第一生成器GB→P的每层编码器结构可以包括LFEB 620、卷积层(例如，Conv 7x7、Conv 3x3等)和激活层(例如，BN+ReLU)，并且每层解码器结构可以包括去卷积层(例如，DeConv 3x3)和激活层(例如，BN+ReLU，Tanh)。残差块RB 610包括卷积层(例如，Conv 3x3)、激活层(例如，BN+ReLU、BN)、丢弃层(例如，Dropout)。
具体地，线能量特征增强块LFEB 620，可以用于实现如上参考步骤S202描述的线能量特征增强处理，如上式(2)所示，其可以包括：
621，对于每个通道的特征图，减去该特征图的平均值，以获得该特征图中的高频分量特征；
622，通过基于高斯-MFRAT核的扩张卷积，从高频分量特征中提取线能量特征；
623，通过最大响应操作，从这些线能量特征中获得该特征图的线能量特征；
具体地，将提取出的多个线能量特征中的最大线能量特征，作为该特征图的线能量特征；
624，将该特征图的线能量特征与预设的学习参数S相乘，并将乘积添加到原始特征图(即增强前的特征图)，从而得到线能量特征增强后的特征图。
本申请实施例中，线能量特征增强块LFEB 620是一种轻量级的即插即用块，可以方便域传输以及提高识别性能。
此外，如图6(a)所示，在联合训练中的第一生成阶段中，还包括基于真实掌纹图像603，获得真实PCE图像604，以用于监督训练PCE图像样本602。
根据本申请实施例，在所述联合训练中，利用掌纹线能量提取器(PCE Extractor,PCEE)，从真实掌纹图像中提取出对应的真实PCE图像。
在本申请实施例中，可以设计用于从真实掌纹图像中提取PCE图像的PCEE，以生成用于联合训练的监督信息，使得所生成的掌纹图像更逼真。图7是示出根据本申请实施例的掌纹线能量提取器PCEE的示意图。如图7所示，与以上关于线能量特征增强块的描述类似地，本申请掌纹线能量提取器PCEE可以包括均值滤波器710、线状卷积层720、最大响应操作层730和自适应二值化层740，以真实掌纹图像700为输入，最终生成二值化的PCE图像750。
根据本申请实施例，利用掌纹线能量提取器，从所述真实掌纹图像中提取出真实掌纹线能量图，可以包括：
利用所述掌纹线能量提取器，提取所述真实掌纹图像中的真实掌纹线，并确定所述真实掌纹线中每个像素的线能量特征；
基于所述真实掌纹线和所述真实掌纹线中每个像素的线能量特征，通过二值化处理，获得所述真实掌纹线能量图。
可选地，如图7所示，应用均值滤波器710，从原始的真实掌纹图像700中减去滤波结果，以获得高频分量；通过线状卷积层720，获得该真实掌纹图像700中的线能量特征，即真实掌纹线的线条分布，其中，该线状卷积层可以由沿着不同方向的若干高斯-MFRAT核组成。
可选地，可以通过最大响应操作730，针对真实掌纹线中的每个像素，基于该像素沿着各个高斯-MFRAT核所对应的各个方向上的线方向能量，确定其中最大的线方向能量作为该像素的线能量特征。
可选地，可以使用自适应二值化处理740来获得最终的PCE图像。例如，为了突出显示主要线条，可以根据整个图像上的线能量特征的值的前10％，设置二值化阈值T。
根据本申请实施例，在联合训练中，所述第二生成器可以以真实掌纹图像的真实掌纹线能量图为输入，以与所述真实掌纹图像相对应的拟真掌纹图像为输出，所述拟真掌纹图像包括细节纹理信息。
图8是示出根据本申请实施例的联合训练中的第二生成阶段的示意图。
如图8所示，联合训练中的第二生成阶段利用真实掌纹图像801参与训练，其中，与上文类似地，同样利用PCEE，从真实掌纹图像801中提取真实PCE图像802，例如，从真实掌纹图像A通过PCEE生成真实PCE图像fPCEE(A)，继而可以将真实PCE图像fPCEE(A)输入到待训练的第二生成器GP→ R，获得具有与真实PCE图像fPCEE(A)一致的掌纹线的拟真掌纹图像A*803。
可选地，在该联合训练中的第二生成阶段中，可以利用编码器E，通过重新参数化方法，将输入的真实掌纹图像A映射到具有均值μQ和方差的隐空间Q(z|A)，该隐空间可以服从正态分布从而基于该隐空间生成用于控制纹理信息生成的隐向量(即，控制向量)。
可选地，可以在训练阶段对该隐向量的散度进行约束，例如，可以使隐空间的分布接近标准正态分布N(0,1)，即，将隐空间近似到标准正态空间，从而容易地将随机噪声z～N(0,1)采样为隐向量，以生成多样化的掌纹细节。
此外，考虑到直接用少量样本进行训练(few-shot training)可能导致鉴别器出现过拟合问题，在本申请实施例中，可以在将所生成的拟真掌纹图像803和真实掌纹图像801馈送到鉴别器D之前，采用数据增强模块AUG对这些图像进行扩充，以生成更多样化的训练样本，从而提高鉴别器D的泛化能力和鲁棒性，并且即使只有少量样本进行训练，也能够减少鉴别器D的过拟合风险，使其更好地适应不同的掌纹图像样本。
根据本申请实施例，在所述联合训练中，可以基于真实PCE图像,对所述第一生成器和所述第二生成器的生成过程进行监督，其中，所述掌纹线能量提取器可以与所述第一生成器和所述第二生成器联合训练。
可选地，如图6(a)所示，真实PCE图像604可以用于监督PCE图像样本602的生成，例如，可以使用从真实掌纹图像603中提取的真实PCE图像604作为对抗样本，以采用对抗损失来驱动第一生成器生成的结果更接近真实PCE图像604，从而优化PCE图像的生成质量。
可选地，如图8所示，可以在联合训练中的第二生成阶段中采用具有PCEE的循环结构，将所生成的拟真掌纹图像A*重新映射回PCE域fPCEE(A*)，即得到拟真PCE图像804，并通过最小化真实PCE图像803与拟真PCE图像804之间的差距，保证拟真掌纹图像与原始的真实掌纹图像之间具有一致的身份信息。例如，可以通过最小化fPCEE(A)与所生成的fPCEE(A*)之间的L1距离来约束所生成的拟真掌纹图像A*的失真，以严格保留输入的fPCEE(A)的ID信息，并且可以应用鉴别器D来强化所生成的拟真掌纹图像A*的真实性。
根据本申请实施例，所述联合训练的损失函数可以包括与所述第一生成器相关的第一损失函数以及与所述第二生成器相关的第二损失函数；
其中，所述第一损失函数包括：用于保持所述掌纹线能量图样本与所述贝塞尔曲线样本之间结构一致性的对比损失、以及用于使所述掌纹线能量图样本与所述真实掌纹线能量图之间相似的对抗损失；
所述第二损失函数包括：用于使控制向量的分布接近标准正态分布的分布控制损失、所述拟真掌纹图像的身份与所述真实掌纹图像的身份之间的身份一致性损失、所述拟真掌纹图像的失真损失、以及强化所述拟真掌纹图像的真实性的损失。
如上所述，本申请联合训练可以包括对第一生成器、第二生成器和PCEE的联合训练，其中，PCEE作为第一生成器与第二生成器之间的连接。因此，整个训练过程的损失函数可以包括与第一生成器的生成过程和第二生成器的生成过程相对应的损失函数。
可选地，如图6(a)所示，与第一生成器的生成过程相对应的损失函数可以包括用于保持结构一致性的对比损失LCL、以及用于使掌纹图像更逼真的对抗损失Ladv。
例如，如图6(a)所示，为了限制贝塞尔曲线样本B和PCE图像样本GB→P(B)之间的结构一致性， 可以使用对比损失(contrastive loss)。具体地，可以从B和GB→P(B)两者中裁剪一系列图块。对于GB→ P(B)中的查询图块q，B中在相同位置处的对应图块可以作为该查询图块q的正样本k+，并且B中的其他图块可以作为该查询图块的负样本因此，可以通过使正样本之间更靠近并且负样本之间更远离的方式来构造对比损失函数，例如，应用归一化互信息神经估计(InfoNCE)损失函数的对比损失LCL可以表示如下：
其中，τ是温度超参数。
因此，与第一生成器的生成过程相对应的损失函数可以表示如下：
其中，fPCEE表示对图像A执行PCEE处理，A表示随机的真实掌纹图像，并且为两个权重。
可选地，如图8所示，与第二生成器的生成过程相对应的损失函数可以包括用于使控制向量N(z)的分布接近标准正态分布Q(z|A)的分布控制损失LKL、用于保持第二生成器基于真实PCE图像802生成的拟真掌纹图像803的身份与真实掌纹图像801的身份之间的身份一致性损失Lcyc、以及用于约束第二生成器生成的拟真掌纹图像803的失真并强化真实性的生成损失LG。
可选地，分布控制损失LKL可以表示如下：
可选地，身份一致性损失Lcyc可以表示如下：
可选地，生成损失LG可以包括针对拟真掌纹图像803的失真损失以及针对鉴别器来强化拟真掌纹图像803的真实性的损失其中，T()表示数据增强模块T的处理。因此，生成损失LG可以表示如下：
因此，与第二生成器的生成过程相对应的损失函数可以表示如下：
其中，表示不同损失项的权重。
当然，以上给出的损失函数的计算方式在本申请中仅用作示例而非限制，本申请同样可以采用其他形式的损失函数。
因此，通过对上述整个训练过程的损失函数的优化，可以确定上述第一生成器、第二生成器以及PCEE中的所有参数，从而应用于本申请掌纹图像生成过程。
下面，将参考图9和图10来呈现对本申请掌纹图像生成方法的性能验证。
可选地，在该性能验证中，可以遵循与贝塞尔掌纹生成方法和RPG-Palm(Realistic Pseudo-data Generation-Palm)掌纹生成方法中相同的实验数据集和开放集评估方案。例如，可以根据TAR、FAR，评估基于各种掌纹生成方法生成的掌纹图像预训练的识别模型的性能，其中，TAR和FAR分别代表“正确接受率(True Acceptance Rate)”和“错误接受率(False Acceptance Rate)”，也就是说，TAR可以表示识别正确的样本被正确接受的比例，也可以理解为识别模型的准确率，而FAR识别错误的样本被错误接受的比例，也可以理解为识别模型的误识率。此外，FID(Fréchet Inception Distance，弗雷歇Inception距离)度量可以用于评估所生成的掌纹图像的质量。
可选地，可以遵循与贝塞尔掌纹生成方法和RPG掌纹生成方法，在该性能验证中采用13个公共数据集，其可以来自各种设备，总共具有3,268个ID和59,162个图像。其中，可以遵循检测后裁剪(detect-then-crop)方案来提取感兴趣区域(ROI)。
可选地，在该性能验证中，可以按照贝塞尔掌纹生成方法和RPG掌纹生成方法，生成4000个身 份，每个身份默认有100个样本。对于第一生成阶段，和τ可以被设置为1.0、1.0和1.0，并且学习速率在前30个训练周期(epoch)中为0.0002，并且在最后30个训练周期中线性衰减到1e-6。对于第二生成阶段，可以将和分别设置为1.0、10.0、0.01和1.0，并且学习速率在前50个训练周期中为0.0002，并且在最后50个训练周期中线性衰减到1e-8。在联合训练阶段，自适应矩估计(Adaptive Moment Estimation，Adam)优化器参数被设置为(0.5,0.99)。上述训练中所有图像的分辨率可以被设置为256×256。
此外，为了实现公平的性能比较，可以采用与贝塞尔掌纹生成方法所对应的掌纹识别模型相同的识别模型骨架，即ResNet(残差网络)50和MobileFaceNet(移动人脸识别网络)，输入图像的分辨率为224×224。该识别模型首先在25个训练周期内在合成数据上进行预训练，然后在50个训练周期内在真实数据集上进行微调。在50个训练周期内在真实数据集上训练比较的基线模型。具有裕度(margin)m＝0.5和比例因子s＝48的ArcFace(弧度面部识别方法)被用于预训练、微调和基线训练监督，预训练和微调的最大和最小学习速率分别可以被设置为1e-2和1e-6。可以使用小批量随机梯度下降(SGD)方法训练所有识别模型，其中批大小可以为128。
因此，基于以上实验设置，在本申请中，首先可以验证在训练身份和测试身份完全隔离的开放集协议下识别模型的性能。可选地，可以采用两种不同的训练ID和测试ID的比率，例如1:1和1:3(例如，训练:测试为1634:1632和818:2448)，定量结果示于下表1中，其中，“MB”表示MobileFaceNet，并且“R50”表示ResNet50。
表1开放集协议下的定量结果
如表1所示，本申请实施例中所述方法可以以清晰的裕度改进RPG掌纹生成方法，并且在训练ID和测试ID的比率分别为1:1和1:3设置下都实现了最高性能水平。此外，本申请实施例中所述方法在训练ID和测试ID的比率为1:3的设置下的改进大于比率为1:1的设置，也就是说，本申请实施例中所述方法在较少真实数据的情况下具有显著有效性。
此外，为了验证本申请实施例中所述方法在有限训练身份下的性能，可以在训练ID和测试ID的比率为1:1的开放集协议下测试具有不同数量的训练身份(ID)的模型。具体地，在该验证过程中合成了共计4000个伪ID，并且每个ID包含100个伪掌纹。采用相同的MobileFaceNet作为不同方法的识别模型骨架，定量结果示于下表2中。
表2不同数量的真实训练身份下的性能
如表2所示，当用非常少的真实ID进行训练时，ArcFace、贝塞尔掌纹生成和RPG掌纹生成方法都变得不可用，而本申请实施例中所述方法仍然能够保持性能。在仅利用2.5％的真实ID(即，40)训练时，本申请实施例中所述方法仍然优于用100％的真实ID(即，1600)训练的ArcFace的结果。在仅利用1％的真实ID进行训练的情况下，本申请实施例中所述方法的TAR仍然与利用50％的真实ID(即，800)的ArcFace训练的TAR相当。
可选地，可以使用四种生成方法pix2pixHD、CycleGAN、BicycleGAN和RPG-Palm进行掌纹生成性能比较，它们都使用40个真实ID与根据RPG-Palm的未配对数据进行重新训练。定量结果示于表3中。
表3在开放集协议下使用不同掌纹生成方法的定量识别结果
如上表3所示，基于本申请实施例中所述方法预训练的掌纹识别模型优于其他方法。此外，本申请实施例中所述方法可以实现40.3的FID评分，显示出显著的改善。
图9是示出根据本申请实施例的利用不同掌纹生成方法的掌纹生成结果的对比图。其中，(a)对应于贝塞尔掌纹生成方法Bézier palm，(b)对应于pix2pixHD方法，(c)对应于CycleGAN方法，(d)对应于BicycleGAN方法，(e)对应于RPG-Palm方法，(f)对应于PCE图像，(g)-(j)对应于本申请实施例中所述方法生成的多样化掌纹图像。
如图9所示，在少量训练数据下，RPG-Palm生成的掌纹图像表现出严重的模糊和不一致的线，而本申请实施例中所述方法仍然保持整体清晰度和ID一致性。另外，本申请实施例中所述方法可以恢复关于掌纹线的更好的细节信息，包括诸如厚度的变化和多个皱褶的交织，而不仅仅是迁移贝塞尔曲线。此外，pix2pixHD、CycleGAN和BicycleGAN方法的结果表现出比RPG-Palm更严重的模糊问题。
此外，为了进一步增强输入掌纹的线能量特征，本申请掌纹图像生成方法还可以将所提出的LFEB应用于掌纹识别模型，例如在掌纹识别模型的骨架的第一卷积层之前并入即插即用的LFEB，以增强输入掌纹图像的线能量特征。
该性能验证还可以包括对消融(ablation)的研究。其中，本申请实施例中所述方法的主要组件可以包括PCEE、用于少样本训练的数据扩增(DA)模块、改进的贝塞尔曲线合成、具有LFEB的生成模型、以及具有LFEB的识别模型，其可以分别使用“P”、“A”、“I”、“G+L”和“R+L”来表示。对于基线生成模型，可以使用两阶段训练方法并去除上述组件。因此，可选地，可以使用40个ID训练消融实验，并且测试集在训练ID和测试ID的比率为1:1的开放集协议下固定。消融实验的结果示于下表4中。
表4不同组件的消融
如表4所示，具有PCEE的模型在13.81％@FAR＝1e-6(即，FAR＝1e-6时，TAR为13.81％)下实现了最大的性能改进，这反映了具有PCEE的模型在有限数据下生成逼真的掌纹样本的优越性。通过与基线生成和识别模型进行比较，LFEB模块通过增强掌纹线能量特征带来了6％@FAR＝1e-6的显著且一致的性能增益。DA模块通过用少量训练样本有效地扩展类内多样性来实现5.52％@FAR＝1e-6的改进。此外，对贝塞尔曲线的改进还通过引入更合理的掌纹线分布而实现更好的性能。
最后，在性能验证中，还可以包括针对LFEB的有效性验证。图10是示出根据本申请实施例的针对线能量特征增强块的有效性验证的结果图。
图10中可视化了在具有和不具有LFEB的情况下MobileFaceNet中层1块的特征，其中，图10(a)为输入的掌纹图像，图10(b)为经过LFEB的掌纹图像，图10(c)对应于不具有LFEB的情况下的特征可视化，图10(d)对应于具有LFEB的情况下的特征可视化。因此，可以看出，模型在不具有LFEB的情况下关注于掌纹线和非线条区域两者，而通过添加LFEB组件，模型可以偏向于关注于掌纹线。
如上所述，通过本申请掌纹图像生成方法，可以利用从预先确定的掌纹曲线模板中确定的控制点来生成贝塞尔曲线，并将该贝塞尔曲线转换为具有皱褶信息的掌纹线能量图，其中该皱褶信息包括与贝塞尔曲线相同的线条分布但不同的线型，其线型由每个像素的线能量特征确定，继而进一步基于具有该皱褶信息的掌纹线能量图生成具有纹理信息的掌纹图像，该掌纹图像的掌纹线与掌纹线能量图的皱褶信息一致，从而生成具有逼真皱褶和纹理的拟真掌纹图像。其中，通过引入掌纹线能量域作为连接贝塞尔掌纹域和掌纹图像域的中间域，避免从贝塞尔曲线直接生成具有皱褶信息和纹理信息的掌纹图像，降低了掌纹图像的生成难度，并且在从掌纹线能量图生成掌纹图像的过程中，在保证掌纹图像具有一致掌纹线的同时生成细节纹理信息，使得能够在保留相同的身份信息的情况下生成具有多样化纹理的拟真掌纹图像，因此降低了对真实数据的依赖，适用于缺乏大规模掌纹数据集的掌纹识别训练。
图11是示出根据本申请实施例的掌纹图像生成装置1100的示意图。
根据本申请实施例，所述掌纹图像生成装置1100可以包括曲线生成模块1101、皱褶生成模块1102、和纹理生成模块1103。
曲线生成模块1101可以被配置为基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线。可选地，曲线生成模块1101可以执行如上参考步骤S201所描述的操作。
皱褶生成模块1102可以被配置为基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应。可选地，皱褶生成模块1102可以执行如上参考步骤S202所描述的操作。
纹理生成模块1103可以被配置为基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。可选地，纹理生成模块1103可以执行如上参考步骤S203所描述的操作。
可选地，皱褶生成模块1102被配置为，对所述贝塞尔曲线进行特征提取，获得所述贝塞尔曲线的多通道特征图；基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征；对各个像素的所述线能量特征进行增强，生成增强后的多通道特征图；及，基于所述增强后的多通道特征图，生成所述掌纹线能量图。
可选地，皱褶生成模块1102被配置为，对于每个通道特征图，执行以下处理：
利用线状卷积层，从该通道特征图中提取每个像素的线方向能量，其中，所述线状卷积层包括沿着多个预定方向的高斯改进有限拉东变换核，所述线方向能量包括该像素沿着所述多个预定方向上的线方向能量；
根据所述线方向能量，获得每个像素的所述线能量特征，所述线能量特征包括该像素在所述多个预定方向上的最大线方向能量以及与所述最大线方向能量相对应的预定方向。
所述纹理生成模块1103被配置为，使用多个控制向量，基于所述掌纹线能量图，生成具有多个细节纹理信息的所述拟真掌纹图像。
可选地，所述掌纹线能量图中的掌纹线，作为身份信息，用于指示所述拟真掌纹图像的身份。
可选地，所述控制向量为随机噪声向量，并且所述细节纹理信息包括光线信息、阴影信息以及皮肤纹理信息中的一种或多种。
可选地，所述皱褶生成模块1102被配置为，基于所述贝塞尔曲线，利用预先训练的第一生成器，生成所述掌纹线能量图；
所述纹理生成模块1103被配置为，基于所述掌纹线能量图，利用预先训练的第二生成器，生成所述拟真掌纹图像；
其中，所述第一生成器和所述第二生成器是利用真实掌纹图像联合训练的。
可选地，在所述联合训练中，
所述第一生成器以贝塞尔曲线样本为输入，以掌纹线能量图样本为输出，所述掌纹线能量图样本包括皱褶信息；
所述第二生成器以真实掌纹图像的真实掌纹线能量图为输入，以与所述真实掌纹图像相对应的拟真掌纹图像为输出，所述拟真掌纹图像包括细节纹理信息；
利用掌纹线能量提取器，从所述真实掌纹图像中提取出所述真实掌纹线能量图；
基于所述真实掌纹线能量图，对所述第一生成器和所述第二生成器进行监督；
所述掌纹线能量提取器与所述第一生成器和所述第二生成器联合训练。
可选地，所述利用掌纹线能量提取器，从所述真实掌纹图像中提取出所述真实掌纹线能量图，包括：
利用所述掌纹线能量提取器，提取所述真实掌纹图像中的真实掌纹线，并确定所述真实掌纹线中每个像素的线能量特征；
基于所述真实掌纹线和所述真实掌纹线中每个像素的线能量特征，通过二值化处理，获得所述真实掌纹线能量图。
可选地，所述第一生成器包括线能量特征增强块，用于对所述贝塞尔曲线样本的多通道特征图的线能量特征进行增强。
可选地，所述联合训练的损失函数包括与所述第一生成器相关的第一损失函数以及与所述第二生成器相关的第二损失函数；
其中，所述第一损失函数包括：用于保持所述掌纹线能量图样本与所述贝塞尔曲线样本之间结构一致性的对比损失、以及用于使所述掌纹线能量图样本与所述真实掌纹线能量图之间相似的对抗损失；
所述第二损失函数包括：用于使控制向量的分布接近标准正态分布的分布控制损失、所述拟真掌纹图像的身份与所述真实掌纹图像的身份之间的身份一致性损失、所述拟真掌纹图像的失真损失、以及强化所述拟真掌纹图像的真实性的损失。
可选地，所述掌纹曲线模板是基于人类真实掌纹线得到的统计信息预先确定的；
其中，所述曲线生成模块1101被配置为，基于所述掌纹曲线模板，确定生成所述多个控制点的区域范围；在所述区域范围内进行采样，得到所述多个控制点。
根据本申请又一方面，还提供了一种电子设备。图12示出了根据本申请实施例的电子设备2000的示意图。
如图12所示，所述电子设备2000可以包括一个或多个处理器2010，和一个或多个存储器2020。其中，所述存储器2020中存储有计算机可读代码，所述计算机可读代码当由所述一个或多个处理器2010运行时，可以执行如上所述的掌纹图像生成方法。
本申请实施例中的处理器可以是一种集成电路芯片，具有信号的处理能力。上述处理器可以是通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等，可以是X86架构或ARM架构的。
一般而言，本申请各种示例实施例可以在硬件或专用电路、软件、固件、逻辑，或其任何组合中实施。某些方面可以在硬件中实施，而其他方面可以在可以由控制器、微处理器或其他计算设备执行的固件或软件中实施。当本申请实施例的各方面被图示或描述为框图、流程图或使用某些其他图形表示时，将理解此处描述的方框、装置、系统、技术或方法可以作为非限制性的示例在硬件、软件、固件、专用电路或逻辑、通用硬件或控制器或其他计算设备，或其某些组合中实施。
例如，根据本申请实施例的方法或装置也可以借助于图13所示的计算设备3000的架构来实现。如图13所示，计算设备3000可以包括总线3010、一个或多个CPU 3020、只读存储器(ROM)3030、随机存取存储器(RAM)3040、连接到网络的通信端口3050、输入/输出组件3060、硬盘3070等。计算设备3000中的存储设备，例如ROM 3030或硬盘3070可以存储本申请提供的掌纹图像生成方法的处理和/或通信使用的各种数据或文件以及CPU所执行的程序指令。计算设备3000还可以包括用户界面3080。当然，图13所示的架构只是示例性的，在实现不同的设备时，根据实际需要，可以省略图13示出的计算设备中的一个或多个组件。
根据本申请又一方面，还提供了一种计算机可读存储介质。所述计算机存储介质上存储有计算机可读指令。当所述计算机可读指令由处理器运行时，可以执行参照以上附图描述的根据本申请实施例的掌纹图像生成方法。本申请实施例中的计算机可读存储介质可以是易失性存储器或非易失性存储器，或可包括易失性和非易失性存储器两者。非易失性存储器可以是只读存储器(ROM)、可编程只读存储器(PROM)、可擦除可编程只读存储器(EPROM)、电可擦除可编程只读存储器(EEPROM)或闪存。易失性存储器可以是随机存取存储器(RAM)，其用作外部高速缓存。通过示例性但不是限制性说明，许多形式的RAM可用，例如静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、同步动态随机存取存储器(SDRAM)、双倍数据速率同步动态随机存取存储器(DDRSDRAM)、增强型同步动态随机存取存储器(ESDRAM)、同步连接动态随机存取存储器(SLDRAM)和直接内存总线随机存取存储器(DR RAM)。应注意，本文描述的方法的存储器旨在包括但不限于这些和任意其它适合类型的存储器。应注意，本文描述的方法的存储器旨在包括但不限于这些和任意其它适合类型的存储器。
本申请实施例还提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行根据本申请实施例的掌纹图像生成方法。
本申请实施例提供了一种掌纹图像生成方法、装置、设备和计算机可读存储介质。
本申请实施例所提供的方法相比于传统的掌纹生成方法而言，能够通过引入掌纹线能量域来对掌纹的皱褶和纹理的生成进行解耦，以通过将贝塞尔掌纹域的贝塞尔曲线转换至掌纹线能量域的掌纹线 能量图来生成逼真的皱褶，并通过将掌纹线能量域的掌纹线能量图转换至掌纹图像域的掌纹图像来生成逼真的纹理，从而降低了从贝塞尔曲线生成拟真掌纹图像的难度，实现了多样化的拟真掌纹图像生成。
本申请实施例所提供的方法利用从预先确定的掌纹曲线模板中确定的控制点来生成贝塞尔曲线，并将该贝塞尔曲线转换为具有皱褶信息的掌纹线能量图，其中该皱褶信息包括与贝塞尔曲线相同的线条分布但不同的线型，其线型由每个像素的线能量特征确定，继而进一步基于具有该皱褶信息的掌纹线能量图生成具有纹理信息的掌纹图像，该掌纹图像的掌纹线与掌纹线能量图的皱褶信息一致，从而生成具有逼真皱褶和纹理的拟真掌纹图像。本申请实施例所提供的方法通过引入掌纹线能量域作为连接贝塞尔掌纹域和掌纹图像域的中间域，避免从贝塞尔曲线直接生成具有皱褶信息和纹理信息的掌纹图像，降低了掌纹图像的生成难度，并且在从掌纹线能量图生成掌纹图像的过程中，在保证掌纹图像具有一致掌纹线的同时生成细节纹理信息，使得能够在保留相同的身份信息的情况下生成具有多样化纹理的拟真掌纹图像，因此降低了对真实数据的依赖，适用于缺乏大规模掌纹数据集的掌纹识别训练。
需要说明的是，附图中的流程图和框图，图示了按照本申请各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，所述模块、程序段、或代码的一部分包含至少一个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
一般而言，本申请各种示例实施例可以在硬件或专用电路、软件、固件、逻辑，或其任何组合中实施。某些方面可以在硬件中实施，而其他方面可以在可以由控制器、微处理器或其他计算设备执行的固件或软件中实施。当本申请实施例的各方面被图示或描述为框图、流程图或使用某些其他图形表示时，将理解此处描述的方框、装置、系统、技术或方法可以作为非限制性的示例在硬件、软件、固件、专用电路或逻辑、通用硬件或控制器或其他计算设备，或其某些组合中实施。
在上面详细描述的本申请示例实施例仅仅是说明性的，而不是限制性的。本领域技术人员应该理解，在不脱离本申请原理和精神的情况下，可对这些实施例或其特征进行各种修改和组合，这样的修改应落入本申请范围内。

Claims (20)

  1. 一种掌纹图像生成方法，由电子设备执行，包括：
    基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线；
    基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应；及，
    基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。
  2. 如权利要求1所述的方法，其中，所述基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，包括：
    对所述贝塞尔曲线进行特征提取，获得所述贝塞尔曲线的多通道特征图；
    基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征；
    对各个像素的所述线能量特征进行增强，生成增强后的多通道特征图；及
    基于所述增强后的多通道特征图，生成所述掌纹线能量图。
  3. 如权利要求2所述的方法，其中，所述基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征，包括：
    对于每个通道特征图，执行以下处理：
    利用线状卷积层，从该通道特征图中提取每个像素的线方向能量，其中，所述线状卷积层包括沿着多个预定方向的高斯改进有限拉东变换核，所述线方向能量包括该像素沿着所述多个预定方向上的线方向能量；
    根据所述线方向能量，获得每个像素的所述线能量特征，所述线能量特征包括该像素在所述多个预定方向上的最大线方向能量以及与所述最大线方向能量相对应的预定方向。
  4. 如权利要求1-3中任一项所述的方法，其中，所述基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像，包括：
    使用多个控制向量，基于所述掌纹线能量图，生成具有多个细节纹理信息的所述拟真掌纹图像。
  5. 如权利要求4所述的方法，其中，所述掌纹线能量图中的掌纹线，作为身份信息，用于指示所述拟真掌纹图像的身份。
  6. 如权利要求4或者5所述的方法，其中，所述控制向量为随机噪声向量，并且所述细节纹理信息包括光线信息、阴影信息以及皮肤纹理信息中的一种或多种。
  7. 如权利要求4-6中任一项所述的方法，其中，所述基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，包括：
    基于所述贝塞尔曲线，利用预先训练的第一生成器，生成所述掌纹线能量图；
    所述基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像，包括：
    基于所述掌纹线能量图，利用预先训练的第二生成器，生成所述拟真掌纹图像；
    其中，所述第一生成器和所述第二生成器是利用真实掌纹图像联合训练的。
  8. 如权利要求7所述的方法，其中，在所述联合训练中，
    所述第一生成器以贝塞尔曲线样本为输入，以掌纹线能量图样本为输出，所述掌纹线能量图样本包括皱褶信息；
    所述第二生成器以真实掌纹图像的真实掌纹线能量图为输入，以与所述真实掌纹图像相对应的拟真掌纹图像为输出，所述拟真掌纹图像包括细节纹理信息；
    利用掌纹线能量提取器，从所述真实掌纹图像中提取出所述真实掌纹线能量图；
    基于所述真实掌纹线能量图，对所述第一生成器和所述第二生成器进行监督；及，
    所述掌纹线能量提取器与所述第一生成器和所述第二生成器联合训练。
  9. 如权利要求8所述的方法，其中，所述利用掌纹线能量提取器，从所述真实掌纹图像中提取出所述真实掌纹线能量图，包括：
    利用所述掌纹线能量提取器，提取所述真实掌纹图像中的真实掌纹线，并确定所述真实掌纹线中每个像素的线能量特征；
    基于所述真实掌纹线和所述真实掌纹线中每个像素的线能量特征，通过二值化处理，获得所述真实掌纹线能量图。
  10. 如权利要求7-9中任一项所述的方法，其中，所述第一生成器包括线能量特征增强块，用于对所述贝塞尔曲线样本的多通道特征图的线能量特征进行增强。
  11. 如权利要求8-10中任一项所述的方法，其中，所述联合训练的损失函数包括与所述第一生成器相关的第一损失函数以及与所述第二生成器相关的第二损失函数；
    其中，所述第一损失函数包括：用于保持所述掌纹线能量图样本与所述贝塞尔曲线样本之间结构一致性的对比损失、以及用于使所述掌纹线能量图样本与所述真实掌纹线能量图之间相似的对抗损失；
    所述第二损失函数包括：用于使控制向量的分布接近标准正态分布的分布控制损失、所述拟真掌纹图像的身份与所述真实掌纹图像的身份之间的身份一致性损失、所述拟真掌纹图像的失真损失、以及强化所述拟真掌纹图像的真实性的损失。
  12. 如权利要求1-11中任一项所述的方法，其中，所述掌纹曲线模板是基于人类真实掌纹线得到的统计信息预先确定的；
    其中，所述基于预先确定的掌纹曲线模板，确定多个控制点，包括：
    基于所述掌纹曲线模板，确定生成所述多个控制点的区域范围；
    在所述区域范围内进行采样，得到所述多个控制点。
  13. 一种掌纹图像生成装置，包括：
    曲线生成模块，被配置为基于预先确定的掌纹曲线模板，确定多个控制点，并基于所述多个控制点生成贝塞尔曲线；
    皱褶生成模块，被配置为基于所述贝塞尔曲线，生成具有皱褶信息的掌纹线能量图，所述掌纹线能量图包括与所述贝塞尔曲线具有相同线条分布、但不同线型的掌纹线，其中，所述线型用于描述所述皱褶信息、并与所述掌纹线上各个像素的线方向能量相对应；及，
    纹理生成模块，被配置为基于所述掌纹线能量图，生成具有细节纹理信息的拟真掌纹图像。
  14. 如权利要求13所述的装置，其中，所述皱褶生成模块被配置为，对所述贝塞尔曲线进行特征提取，获得所述贝塞尔曲线的多通道特征图；基于所述多通道特征图，确定每个通道特征图中每个像素的线能量特征；对各个像素的所述线能量特征进行增强，生成增强后的多通道特征图；及，基于所述增强后的多通道特征图，生成所述掌纹线能量图。
  15. 如权利要求13或14所述的装置，其中，所述纹理生成模块被配置为，使用多个控制向量，基于所述掌纹线能量图，生成具有多个细节纹理信息的所述拟真掌纹图像。
  16. 如权利要求13所述的装置，其中，所述皱褶生成模块被配置为，基于所述贝塞尔曲线，利用预先训练的第一生成器，生成所述掌纹线能量图；
    所述纹理生成模块被配置为，基于所述掌纹线能量图，利用预先训练的第二生成器，生成所述拟真掌纹图像；
    其中，所述第一生成器和所述第二生成器是利用真实掌纹图像联合训练的。
  17. 如权利要求13-16中任一项所述的装置，其中，所述掌纹曲线模板是基于人类真实掌纹线得到的统计信息预先确定的；
    其中，所述曲线生成模块被配置为，基于所述掌纹曲线模板，确定生成所述多个控制点的区域范围；在所述区域范围内进行采样，得到所述多个控制点。
  18. 一种电子设备，包括：
    一个或多个处理器；以及
    一个或多个存储器，其中存储有计算机可执行程序，当由所述处理器执行所述计算机可执行程序时，执行权利要求1-12中任一项所述的方法。
  19. 一种计算机程序产品，所述计算机程序产品存储在计算机可读存储介质上，并且包括计算机指令，所述计算机指令在由处理器运行时，使得计算机设备执行权利要求1-12中任一项所述的方法。
  20. 一种计算机可读存储介质，其上存储有计算机可执行指令，所述指令在被处理器执行时用于实现如权利要求1-12中任一项所述的方法。
PCT/CN2024/123335 2023-11-17 2024-10-08 掌纹图像生成方法、装置、设备、存储介质及程序产品 Pending WO2025103010A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202311546177.7 2023-11-17
CN202311546177.7A CN118608635A (zh) 2023-11-17 2023-11-17 掌纹图像生成方法、装置、设备和存储介质

Publications (1)

Publication Number Publication Date
WO2025103010A1 true WO2025103010A1 (zh) 2025-05-22

Family

ID=92557769

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/123335 Pending WO2025103010A1 (zh) 2023-11-17 2024-10-08 掌纹图像生成方法、装置、设备、存储介质及程序产品

Country Status (2)

Country Link
CN (1) CN118608635A (zh)
WO (1) WO2025103010A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN121167941A (zh) * 2025-11-20 2025-12-19 华北电力大学（保定） 基于主副代理模型的轴流压气机叶型优化方法及相关装置

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118608635A (zh) * 2023-11-17 2024-09-06 腾讯科技（深圳）有限公司 掌纹图像生成方法、装置、设备和存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115527079A (zh) * 2022-02-28 2022-12-27 腾讯科技（深圳）有限公司 掌纹样本的生成方法、装置、设备、介质及程序产品
US20230075233A1 (en) * 2020-01-30 2023-03-09 The Regents Of The University Of California Synthetic human fingerprints
CN116168269A (zh) * 2023-01-19 2023-05-26 支付宝(杭州)信息技术有限公司 掌纹图像生成模型的训练方法及系统、掌纹图像生成方法及系统
CN116994297A (zh) * 2022-09-07 2023-11-03 腾讯科技（深圳）有限公司 掌纹图像生成方法、掌纹识别模型训练方法、装置及介质
CN118608635A (zh) * 2023-11-17 2024-09-06 腾讯科技（深圳）有限公司 掌纹图像生成方法、装置、设备和存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20230075233A1 (en) * 2020-01-30 2023-03-09 The Regents Of The University Of California Synthetic human fingerprints
CN115527079A (zh) * 2022-02-28 2022-12-27 腾讯科技（深圳）有限公司 掌纹样本的生成方法、装置、设备、介质及程序产品
CN116994297A (zh) * 2022-09-07 2023-11-03 腾讯科技（深圳）有限公司 掌纹图像生成方法、掌纹识别模型训练方法、装置及介质
CN116168269A (zh) * 2023-01-19 2023-05-26 支付宝(杭州)信息技术有限公司 掌纹图像生成模型的训练方法及系统、掌纹图像生成方法及系统
CN118608635A (zh) * 2023-11-17 2024-09-06 腾讯科技（深圳）有限公司 掌纹图像生成方法、装置、设备和存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN121167941A (zh) * 2025-11-20 2025-12-19 华北电力大学（保定） 基于主副代理模型的轴流压气机叶型优化方法及相关装置

Also Published As

Publication number Publication date
CN118608635A (zh) 2024-09-06

Similar Documents

Publication Publication Date Title
Zhao et al. Dermoscopy image classification based on StyleGAN and DenseNet201
WO2025103010A1 (zh) 掌纹图像生成方法、装置、设备、存储介质及程序产品
WO2020063527A1 (zh) 基于多特征检索和形变的人体发型生成方法
CN110569756A (zh) 人脸识别模型构建方法、识别方法、设备和存储介质
CN110348319A (zh) 一种基于人脸深度信息和边缘图像融合的人脸防伪方法
CN109949255A (zh) 图像重建方法及设备
CN110348330A (zh) 基于vae-acgan的人脸姿态虚拟视图生成方法
CN107680119A (zh) 一种基于时空上下文融合多特征及尺度滤波的跟踪算法
CN109002763B (zh) 基于同源连续性的模拟人脸老化的方法及装置
CN105354555B (zh) 一种基于概率图模型的三维人脸识别方法
CN113538221A (zh) 三维人脸的处理方法、训练方法、生成方法、装置及设备
CN114882545B (zh) 基于三维智能重建的多角度人脸识别方法
CN113327191A (zh) 人脸图像合成方法及装置
CN110414516A (zh) 一种基于深度学习的单个汉字识别方法
CN118553001A (zh) 基于素描输入的纹理可控的三维精细人脸重建方法及装置
CN119228938A (zh) 一种基于感知损失的人脸素描合成方法及系统
CN118052700A (zh) 一种基于改进CycleGAN的人脸属性风格转换方法及系统
CN110956116B (zh) 基于卷积神经网络的人脸图像性别识别模型及识别方法
CN119649410B (zh) 一种基于三维重建和图像生成的行人重识别方法
Feng et al. Study on the optimization of CNN based on image identification
CN102214292B (zh) 人脸图像的光照处理方法
CN113705480A (zh) 基于姿态识别神经网络的姿态识别方法、设备和介质
Jia et al. Face aging with improved invertible conditional GANs
CN117315736B (zh) 一种基于数据增广结构的端到端3d人脸识别系统
CN104463190B (zh) 年龄估计方法及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24890368

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2024890368

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 24890368.4

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 11202600985P

Country of ref document: SG

WWP Wipo information: published in national office

Ref document number: 11202600985P

Country of ref document: SG