WO2021243926A1 - 指静脉识别与防伪一体化方法、装置、存储介质和设备 - Google Patents

指静脉识别与防伪一体化方法、装置、存储介质和设备 Download PDF

Info

Publication number
WO2021243926A1
WO2021243926A1 PCT/CN2020/123182 CN2020123182W WO2021243926A1 WO 2021243926 A1 WO2021243926 A1 WO 2021243926A1 CN 2020123182 W CN2020123182 W CN 2020123182W WO 2021243926 A1 WO2021243926 A1 WO 2021243926A1
Authority
WO
WIPO (PCT)
Prior art keywords
counterfeiting
recognition
finger vein
task
finger
Prior art date
Application number
PCT/CN2020/123182
Other languages
English (en)
French (fr)
Inventor
康文雄
杨伟力
Original Assignee
华南理工大学
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 华南理工大学 filed Critical 华南理工大学
Priority to AU2020432845A priority Critical patent/AU2020432845B2/en
Publication of WO2021243926A1 publication Critical patent/WO2021243926A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components

Definitions

  • the present invention relates to the field of image processing technology, and more specifically, to a method, device, storage medium and equipment for integrating finger vein recognition and anti-counterfeiting.
  • Biometric recognition technology is a promising technology that uses the physiological or behavioral characteristics of the human body to identify individuals through feature extraction methods.
  • biometric recognition technologies include fingerprints, face, iris, gait, voiceprint, palmprint, palm vein, finger vein, signature and so on.
  • Finger veins are distributed under the epidermis, which has unique advantages compared with other biometrics in terms of its recognition principle: (1) Finger vein collection is captured by infrared camera, the collection method does not need to be touched, and the user-friendliness is good; (2) Vein imaging The required camera is not demanding, and the acquisition hardware is light, easy to realize commercialization; (3) The finger veins are distributed under the epidermis, which is not easy to be damaged, and the safety is high.
  • the existing research often regards the anti-counterfeiting algorithm and the recognition algorithm as two independent subtasks, which reduces the convenience and real-time performance of the system to a certain extent.
  • the combination of the identification algorithm and the anti-counterfeiting algorithm is still a blank in the prior art.
  • the purpose of the present invention is to provide an integrated method, device, storage medium and equipment for finger vein recognition and anti-counterfeiting; the present invention integrates the two tasks of finger vein recognition and finger vein anti-counterfeiting.
  • the unified algorithm can improve the efficiency of vein recognition and the real-time performance of the system while ensuring the accuracy of recognition and anti-counterfeiting.
  • an integrated method for finger vein recognition and anti-counterfeiting which is characterized in that it includes:
  • the pre-processed finger vein data is input to the finger vein recognition anti-counterfeiting task convolutional neural network model, and the finger vein image is identified and anti-counterfeit processed through the finger vein recognition anti-counterfeiting task convolutional neural network model to obtain anti-counterfeiting Task classification probability p and recognition task feature vector v; wherein the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model;
  • the recognition task feature vector v of the vein data is output and saved as the registration sample recognition task feature vector;
  • the recognition mode by comparing the classification probability p of the anti-counterfeiting task of the vein data with the probability threshold s 1 , and the identification task feature vector v of the vein data and the cosine distance of the identification task feature vector of each registered sample and the distance threshold s 2 Compare and output the judgment result.
  • the finger vein recognition anti-counterfeiting task convolutional neural network model includes a basic recognition network and an anti-counterfeiting branch;
  • the basic recognition network includes two convolutional networks, three convolutional modules, and a fully connected layer connected in sequence;
  • the anti-counterfeiting branch includes a convolution module and two fully connected layers connected in sequence; the front end of the convolution module of the anti-counterfeiting branch is inserted after the first convolution module of the basic recognition network, thereby constructing a single-input and multiple-output finger vein recognition anti-counterfeiting Task convolutional neural network model.
  • each of the three convolution modules of the basic recognition network includes two convolution sub-modules and a maximum pooling layer connected in sequence.
  • the size of the convolution kernel is 3*3, the number of channels is 64, and the step size is 2 and 1, respectively;
  • the number of input channels is 64, 128, and 256 respectively; among the three convolution modules, the convolution kernel of the former convolution submodule is 3*3, and the latter convolution
  • the convolution kernel of the submodule is 1*1, the step size of the convolution submodule is 1, the number of input channels of the convolution submodule is the number of input channels of the corresponding convolution module, and the convolution kernel of the maximum pooling layer is 2*2;
  • the fully connected layer of the basic identification network the number of output channels is 512;
  • the output channels are 16 and 2 respectively.
  • the preprocessing of the finger vein image refers to:
  • the active window summation method is used to obtain the brightness statistical curve trend of the finger axial direction of the original finger vein image; the active window adopts the same height as the original finger vein image, and slides column by column with the width of 1/20 of the original finger vein image. Calculate the pixel sum in the window; the two peaks of the brightness statistical curve trend are set as the two interphalangeal joints of the finger; the ROI is intercepted between the two interphalangeal joints as the preprocessed finger vein data.
  • the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model, which refers to:
  • the center loss is used as the loss function, and the center loss is:
  • N represents the number of samples
  • x represents the feature vector of the recognition task output by the network
  • c represents the center of the category
  • Training evaluation indicators are:
  • ⁇ represents the proportion of the basic recognition network and the anti-counterfeiting branch
  • SEER represents the equal error rate of the anti-counterfeiting branch
  • EER represents the equal error rate of the basic recognition network
  • An integrated device for finger vein recognition and anti-counterfeiting which is characterized in that it comprises:
  • the preprocessing module is used to preprocess finger vein images to obtain preprocessed finger vein data
  • the feature extraction module is used to input the preprocessed finger vein data to the finger vein recognition anti-counterfeiting task convolutional neural network model, and recognize the finger vein image through the finger vein recognition anti-counterfeiting task convolutional neural network model And anti-counterfeiting processing to obtain the anti-counterfeiting task classification probability p and the recognition task feature vector v; wherein the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model ;
  • the registration module is used to implement the registration mode.
  • the identification task feature vector v of the vein data is output and saved as the registration sample identification task feature vector;
  • the recognition module is used to realize the recognition mode by comparing the anti-counterfeiting task classification probability p of the vein data with the probability threshold s 1 , and the cosine distance and distance between the recognition task feature vector v of the vein data and the recognition task feature vector of each registered sample The threshold s 2 is compared, and the judgment result is output.
  • a storage medium wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the above-mentioned integrated method for finger vein recognition and anti-counterfeiting.
  • a computing device includes a processor and a memory for storing an executable program for the processor, wherein the processor implements the above-mentioned integrated method for finger vein recognition and anti-counterfeiting when the processor executes the program stored in the memory.
  • the present invention has the following advantages and beneficial effects:
  • the present invention integrates the two tasks of finger vein recognition and finger vein anti-counterfeiting into a unified algorithm, utilizes the powerful learning and fitting ability of neural network, and guarantees the performance of both based on multi-task learning, which can ensure the accuracy of recognition and anti-counterfeiting at the same time, Improved the efficiency of vein recognition and the real-time performance of the system;
  • the basic recognition network has the characteristics of lightweight, its network layers are small, and the two-dimensional convolution kernel is adopted, which greatly reduces the amount of network parameters ;
  • the training process of the finger vein recognition anti-counterfeiting task convolutional neural network model proposed in the present invention uses multi-task evaluation indicators to screen the finger vein recognition anti-counterfeiting task convolutional neural network model, which can effectively guarantee the finger vein recognition anti-counterfeiting task convolutional nerve The performance of the network model.
  • Figure 1 is a flow chart of the integrated method for finger vein recognition and anti-counterfeiting of the present invention
  • FIG. 2 is a model diagram of the convolutional neural network model of the finger vein recognition anti-counterfeiting task of the present invention
  • Fig. 3 is a finger vein image input by the present invention.
  • FIG. 4 is a schematic diagram of the process of preprocessing finger vein images according to the present invention.
  • Figure 5 (a) is the upper edge extraction operator of the present invention.
  • Figure 5(b) is the edge extraction operator of the present invention.
  • Figure 6 is the real and fake vein images input by the present invention.
  • Fig. 7 is a training schematic diagram of the convolutional neural network model of the finger vein recognition anti-counterfeiting task of the present invention.
  • an integrated method for finger vein recognition and anti-counterfeiting includes:
  • the pre-processed finger vein data is input to the finger vein recognition anti-counterfeiting task convolutional neural network model, and the finger vein image is identified and anti-counterfeit processed through the finger vein recognition anti-counterfeiting task convolutional neural network model to obtain anti-counterfeiting Task classification probability p and recognition task feature vector v; wherein the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model;
  • the recognition task feature vector v of the vein data is output and saved as the registration sample recognition task feature vector;
  • the recognition mode by comparing the anti-counterfeiting task classification probability p of the vein data with the probability threshold s 1 , and the identification task feature vector v of the vein data and the cosine distance and distance threshold s 2 of the identification task feature vector of each registered sample Compare and output the judgment result.
  • the finger vein recognition anti-counterfeiting task convolutional neural network model includes a basic recognition network and an anti-counterfeiting branch; the basic recognition network includes two convolutional networks, three convolutional modules, and full connections connected in sequence
  • the anti-counterfeiting branch includes a convolution module and two fully connected layers connected in sequence; the front end of the convolution module of the anti-counterfeiting branch is inserted after the first convolution module of the basic recognition network, thereby constructing a single-input multiple-output Finger vein recognition anti-counterfeiting task convolutional neural network model.
  • the basic recognition network mainly focuses on extracting vein texture features, and the anti-counterfeiting branch focuses on the background information of vein samples. Because the feature information extracted by the basic recognition network and the anti-counterfeiting branch are different, the front and back positions of the two branch points will have a key impact on the two tasks.
  • the shared feature extraction part When the branch point is more forward, the shared feature extraction part will be reduced, which increases the difficulty of feature extraction of the anti-counterfeiting branch, thereby reducing its anti-counterfeiting performance; while the branch point is back, the shared feature extraction part of the two tasks increases, which will lead to the generation of both tasks Influencing each other, the difference in the focus of feature extraction between the two will lead to a decline in the performance of shared feature extraction, thereby reducing the performance of the two tasks.
  • Each of the three convolution modules of the basic recognition network includes two convolution sub-modules and a maximum pooling layer connected in sequence.
  • the size of the convolution kernel is 3*3, the number of channels is 64, and the step size is 2 and 1, respectively;
  • the number of input channels is 64, 128, and 256 respectively; among the three convolution modules, the convolution kernel of the former convolution submodule is 3*3, and the latter convolution
  • the convolution kernel of the submodule is 1*1, the step size of the convolution submodule is 1, the number of input channels of the convolution submodule is the number of input channels of the corresponding convolution module, and the convolution kernel of the maximum pooling layer is 2*2;
  • the fully connected layer of the basic identification network the number of output channels is 512;
  • the output channels are 16 and 2 respectively.
  • the input finger vein image is shown in Figure 3.
  • the finger vein image used in the present invention uses infrared light to irradiate one side of the finger, and uses an infrared camera to image the other side to obtain the finger vein image after preliminary collection. Then it is necessary to pre-process the finger vein image. This is because there is a certain degree of flexibility in placing the finger on the vein acquisition device, and the vein pattern changes caused by the offset rotation of the finger vein image will also exist. In addition, the collected finger vein images contain most of the background information. This information that has nothing to do with the finger veins will interfere and affect the subsequent recognition to a certain extent. In the preprocessing stage, the ROI interception and rotation correction of the finger image are required. Minimize the influence of finger shift and background noise.
  • the demarcation characteristics of the edge of the finger are mainly embodied as the upper and lower edge features.
  • ordinary first-order differential operators can be used. Because the left and right directions of the fingers are along the axial direction of the fingers, there is no need to consider, so only use The vertical edge template is fine.
  • the feature of the upper edge is that the upper gray value is greater than the lower gray value, while the lower edge has the opposite gray value.
  • the lower gray value is greater than the upper gray value. Therefore, the operator shown in Figure 5(a) can be used as the upper gray value.
  • the edge extraction operator, the operator shown in Figure 5(b) is used as the lower edge extraction operator for extraction.
  • the vertical midpoint set can be calculated from the upper and lower edges of the finger extracted in the previous step, and then the midline of the finger can be fitted by the least square method to obtain The tilt angle of the finger from the horizontal direction. Then rotate the image to the corresponding angle to correct the finger to the horizontal direction. Because the finger veins are mainly concentrated in the center area of the finger, by calculating the lowest point and highest point of the upper and lower edges of the finger after rotation, the vertical intercepting contour line is determined.
  • the collection of finger veins generally collects images of the index finger, middle finger and ring finger. These fingers have two interphalangeal joints. There is less tissue between these phalangeal joints due to the presence of synovial fluid. The absorption rate of infrared light is It is lower, so that the average brightness of the knuckle area is greater than that of other finger areas. Using this feature, the user can locate the intercepted area in the axial direction more stably.
  • the active window summation method is used to obtain the trend of the brightness statistical curve of the finger axis of the original image.
  • the active window adopts the same height as the original image, and slides column by column with 1/20 the width of the original image, and calculates the sum of pixels in the window.
  • Gaussian smoothing is performed on the data first to eliminate the interference of outliers. In the figure, it can be seen that there are two obvious wave crests, which correspond to the two interphalangeal joints of the fingers.
  • this article only uses the position information at the maximum peak, extending to the left to 1/3 of the starting point distance as the left end point, extending 2/3 of the image length to the right, and intercepting horizontally according to this area , Get the final ROI as the pre-processed finger vein data.
  • the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model, which refers to:
  • Figure 6 shows the comparison between fake veins and real veins.
  • the forgery generation method is to select a sample in one of the categories of the identification sample set, print the sample on two laser printing films using a printer, and then stack the two films and align them. Place a piece of high-quality white paper in the aligned two films to make a fake model; then put the fake model into the vein collection model for collection, and finally get a fake sample.
  • the basic recognition network and the anti-counterfeiting branch in the convolutional neural network model of the finger vein recognition anti-counterfeiting task are alternately trained, as shown in Figure 7; in the training process, only One of the basic recognition network and the anti-counterfeiting branch participates in the training, and the weight of the other is fixed; reducing the performance impact of samples between tasks.
  • the center loss is used as the loss function, and the center loss is:
  • N represents the number of samples
  • x represents the feature vector of the recognition task output by the network
  • c represents the center of the category
  • Training evaluation indicators are:
  • ⁇ represents the proportion of the basic recognition network and the anti-counterfeiting branch
  • SEER represents the equal error rate of the anti-counterfeiting branch
  • EER represents the equal error rate of the basic recognition network
  • the method of the present invention was compared with the existing anti-counterfeiting methods. Then, in order to verify the robustness and generalization ability of the algorithm, relevant experiments were carried out on the IDIAP anti-counterfeiting database and the SCUT database. In order to verify the impact of additional identification tasks on anti-counterfeiting tasks, a network of the same structure is trained separately under anti-counterfeiting tasks for comparison and verification. The results are shown in Table 1 below.
  • the anti-counterfeiting task is a binary classification task, and its difficulty is relatively simpler than the recognition task. Therefore, the deep network can handle the combination of the two tasks relatively effectively.
  • the addition of the recognition task has less impact on the anti-counterfeiting task.
  • the proposed algorithm is in the identification and anti-counterfeiting tasks. There are unique advantages in combining tasks.
  • the use of unified identification and anti-counterfeiting indicators can more effectively evaluate the performance of the system.
  • the equal error rate EER of the identification task and the HTER index of the anti-counterfeiting task are combined using weights, and a simplified weight combination method is used as the GEER index as an evaluation index for the combination of anti-counterfeiting and identification.
  • the indicators are shown in Table 3 below.
  • the proposed integrated model of vein recognition and anti-counterfeiting is evaluated in terms of time consumption.
  • the deployment platform of the algorithm is carried out on the JetsonTK1 development board.
  • the realization of the algorithm is mainly based on the C++ language for reproduction.
  • the framework of the deep model uses the Tensorflow framework on the platform.
  • the final time consumption is shown in Table 4 below.
  • the average value of the proposed algorithm model in 100 forward operations is 13.11, and its real-time performance is effectively guaranteed in the application system where the algorithm is deployed.
  • the present invention is compared with a variety of traditional methods and depth methods. After the anti-counterfeiting task is added, its performance results are still relatively competitive. In addition, through simplified anti-counterfeiting and identification task indicators, the performance of the overall identification and anti-counterfeiting integrated system is evaluated, and finally time-consuming experiments are used to prove that the algorithm has high real-time performance.
  • this embodiment provides an integrated finger vein recognition and anti-counterfeiting device, including:
  • the preprocessing module is used to preprocess finger vein images to obtain preprocessed finger vein data
  • the feature extraction module is used to input the preprocessed finger vein data to the finger vein recognition anti-counterfeiting task convolutional neural network model, and recognize the finger vein image through the finger vein recognition anti-counterfeiting task convolutional neural network model And anti-counterfeiting processing to obtain the anti-counterfeiting task classification probability p and the recognition task feature vector v; wherein the finger vein recognition anti-counterfeiting task convolutional neural network model is a model obtained by training the initial finger vein recognition anti-counterfeiting task convolutional neural network model ;
  • the registration module is used to implement the registration mode.
  • the identification task feature vector v of the vein data is output and saved as the registration sample identification task feature vector;
  • the recognition module is used to realize the recognition mode, by comparing the anti-counterfeiting task classification probability p of the vein data with the probability threshold s 1 , and the cosine distance and distance between the recognition task feature vector v of the vein data and the recognition task feature vector of each registered sample The threshold s 2 is compared, and the judgment result is output.
  • the storage medium of this embodiment is characterized in that the storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the integrated finger vein recognition and anti-counterfeiting described in the first embodiment. ⁇ method.
  • a computing device of this embodiment includes a processor and a memory for storing an executable program for the processor.
  • the feature is that when the processor executes the program stored in the memory, it implements the finger vein recognition and anti-counterfeiting described in the first embodiment. Integrated approach.

Abstract

一种指静脉识别与防伪一体化方法、装置、存储介质和设备；方法包括：对指静脉图像进行预处理；将预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，得到防伪任务分类概率p和识别任务特征向量v；在注册模式下，当防伪任务分类概率p≤概率阈值s 1，则将识别任务特征向量v输出并保存；在识别模式下，通过对防伪任务分类概率p与概率阈值s 1进行比较，以及对识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。该方法将指静脉识别和指静脉防伪两个任务整合到统一算法中，可在保证识别和防伪精度的同时，提升了静脉识别效率和系统实时性。

Description

指静脉识别与防伪一体化方法、装置、存储介质和设备 技术领域
本发明涉及图像处理技术领域，更具体地说，涉及指静脉识别与防伪一体化方法、装置、存储介质和设备。
背景技术
生物特征识别技术是一种利用人体的生理或行为特征，通过特征提取方法对个体进行鉴别的前景广阔的技术。目前常用的生物特征识别技术有指纹、人脸、虹膜、步态、声纹、掌纹、掌静脉、指静脉、签名等等。手指静脉分布在表皮之下，从其识别原理上与其他生物识别相比具有独特的优势：(1)指静脉采集通过红外摄像头捕获，采集方式无需接触，用户友好度好；(2)静脉成像需要的摄像头要求不高，而且采集硬件轻型，容易实现产品化；(3)指静脉分布在表皮之下，不易受到损害，安全性较高。
基于以上优点，指静脉识别在科研界和工业界受到越来越多关注，并其应用场景逐渐多样化与普及化。另外，生物特征识别的安全性能引起了人们的讨论与担忧，近年来出现的多种仿冒攻击方法也让生物特征识别受到了挑战，生物特征识别的仿冒检测能力是衡量其系统稳定性的重要指标，为了提升生物特征识别系统的安全性能，针对性的设计防伪检测算法是有效地解决方式。
但已有的研究往往将防伪算法与识别算法作为两个独立的子任务进行研究，在一定程度上降低了系统的便捷性和实时性。而识别算法与防伪算法结合为一体的方式在现有技术上仍是空白。
发明内容
为克服现有技术中的缺点与不足，本发明的目的在于提供一种指静脉识别与防伪一体化方法、装置、存储介质和设备；本发明将指静脉识别和指静脉防伪两个任务整合到统一算法中，可在保证识别和防伪精度的同时，提升了静脉识别效率和系统实时性。
为了达到上述目的，本发明通过下述技术方案予以实现：一种指静脉识别与防伪一体化方法，其特征在于：包括：
获取待识别的指静脉图像；对指静脉图像进行预处理，得到预处理后的指静脉数据；
将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
在注册模式下，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
在识别模式下，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
优选地，所述指静脉识别防伪任务卷积神经网络模型包括基础识别网络和防伪分支；所述基础识别网络包括依次连接的两个卷积网络、三个卷积模块和全连接层；所述防伪分支包括依次连接的一个卷积模块和两个全连接层；防伪分支的卷积模块前端插入于基础识别网络的第一个卷积模块之后，从而构建出单输入多输出的指静脉识别防伪任务卷积神经网络模型。
优选地，所述基础识别网络的三个卷积模块均包括依次连接的两个卷积子模块和最大值池化层。
优选地，所述基础识别网络的两个卷积网络中，卷积核大小均为3*3，通道数均为64，步长分别为2和1；
所述基础识别网络的三个卷积模块中，输入通道数分别为64、128和256；三个卷积模块中，前一卷积子模块的卷积核为3*3，后一卷积子模块的卷积核为1*1，卷积子模块的步长均为1，卷积子模块的输入通道数为对应卷积模块的输入通道数，最大值池化层的卷积核为2*2；
所述基础识别网络的全连接层，输出通道数为512；
所述防伪分支的全连接层中，输出通道分别为16和2。
优选地，所述对指静脉图像进行预处理，是指：
在指静脉图像提取手指上下边缘；提取手指上下边缘的垂直中点集合，通过最小二乘法拟合手指中线，从而求出手指与水平方向的倾斜角；然后对指静脉图像旋转，将手指矫正到水平方向；
采用活动窗口求和的方法来获取原始指静脉图像手指轴向方的亮度统计曲线趋势；活动窗口采用与原始指静脉图像一致的高度，和原始指静脉图像宽度1/20的宽度逐列滑动，计算窗口内的像素和；亮度统计曲线趋势的两个波峰设定为手指的两个指间关节；在两个指间关节之间截取ROI作为预处理后的指静脉数据。
优选地，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型，是指：
训练样本包括识别注册样本集合和防伪样本集合；识别样本集合为R＝{r 1,r 2,...,r n}，防伪样本集合为S＝{s 1,s 2,...,s n}；其中，防伪样本集合中的伪造样本，根据识别样本中的类别进行伪造生成，即s i＝f(r i)，0≤i＜n；
以遍历识别样本集合和防伪样本集合为迭代单位，交替对指静脉识别防伪任务卷积神经网络模型中的基础识别网络和防伪分支进行训练；在训练过程中，每次只有基础识别网络和防伪分支的其中一项参与训练，另一项的权重固定；
在基础识别网络和防伪分支的训练中，均使用中心损失作为损失函数，其中中心损失为：
Figure PCTCN2020123182-appb-000001
其中，N表示样本数目，x代表网络输出的识别任务特征向量，c表示该类别的中心；
训练评价指标为：
GEER ω＝ω·SEER+(1-ω)·EER′
其中，ω表示基础识别网络和防伪分支的比重，SEER表示防伪分支的等误率，EER表示基础识别网络的等误率。
优选地，在所述识别模式下，采用如下两种方式之一：
一、先判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指 静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若存在余弦距离大于距离阈值s 2时，则输出通过结果；否则输出拒绝结果；
二、先将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若不存在余弦距离大于距离阈值s 2时，则输出拒绝结果；否则继续判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则输出通过结果。
一种指静脉识别与防伪一体化装置，其特征在于，包括：
预处理模块，用于对指静脉图像进行预处理，得到预处理后的指静脉数据；
特征提取模块，用于将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
注册模块，用于实现注册模式，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
识别模块，用于实现识别模式，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
一种存储介质，其特征在于，其中所述存储介质存储有计算机程序，所述计算机程序当被处理器执行时使所述处理器执行上述指静脉识别与防伪一体化方法。
一种计算设备，包括处理器以及用于存储处理器可执行程序的存储器，其特征在于，所述处理器执行存储器存储的程序时，实现上述指静脉识别与防伪一体化方法。
与现有技术相比，本发明具有如下优点与有益效果：
1、本发明将指静脉识别和指静脉防伪两个任务整合到统一算法中，利用神 经网络强大的学习拟合能力，基于多任务学习保证两者性能，可在保证识别和防伪精度的同时，提升了静脉识别效率和系统实时性；
2、本发明提出的指静脉识别防伪任务卷积神经网络模型中，基础识别网络具有轻量化的特点，其网络层数较少，且采用了二维卷积核，大大降低了网络的参数量；
3、本发明提出的指静脉识别防伪任务卷积神经网络模型的训练处理，利用多任务评价指标对指静脉识别防伪任务卷积神经网络模型进行筛选，可以有效保证指静脉识别防伪任务卷积神经网络模型的性能。
附图说明
图1是本发明指静脉识别与防伪一体化方法的流程图；
图2是本发明指静脉识别防伪任务卷积神经网络模型的模型图；
图3是本发明输入的指静脉图像；
图4是本发明对指静脉图像进行预处理的过程示意图；
图5(a)是本发明上边缘提取算子；
图5(b)是本发明下边缘提取算子；
图6是本发明输入的真实和伪造静脉图像；
图7是本发明指静脉识别防伪任务卷积神经网络模型的训练示意图。
具体实施方式
下面结合附图与具体实施方式对本发明作进一步详细的描述。
实施例一
本实施例一种指静脉识别与防伪一体化方法，其流程如图1所示，包括：
获取待识别的指静脉图像；对指静脉图像进行预处理，得到预处理后的指静脉数据；
将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
在注册模式下，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
在识别模式下，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
所述指静脉识别防伪任务卷积神经网络模型，如图2所示，包括基础识别网络和防伪分支；所述基础识别网络包括依次连接的两个卷积网络、三个卷积模块和全连接层；所述防伪分支包括依次连接的一个卷积模块和两个全连接层；防伪分支的卷积模块前端插入于基础识别网络的第一个卷积模块之后，从而构建出单输入多输出的指静脉识别防伪任务卷积神经网络模型。
在指静脉识别防伪任务卷积神经网络模型中，基础识别网络主要聚焦于提取静脉纹理特征，防伪分支关注静脉样本背景信息。因为基础识别网络和防伪分支提取的特征信息有所差别，两者分支点的前后位置将对两种任务造成关键影响。当分支点更靠前时，共有特征提取部分将缩减，增加了防伪分支的特征提取难度，从而降低其防伪性能；而分支点靠后时，两任务共享特征提取部分增加，将导致两者任务产生相互影响，两者特征提取侧重的差异性将导致共有特征提取性能下降，从而降低两者任务的性能。
所述基础识别网络的三个卷积模块均包括依次连接的两个卷积子模块和最大值池化层。
所述基础识别网络的两个卷积网络中，卷积核大小均为3*3，通道数均为64，步长分别为2和1；
所述基础识别网络的三个卷积模块中，输入通道数分别为64、128和256；三个卷积模块中，前一卷积子模块的卷积核为3*3，后一卷积子模块的卷积核为1*1，卷积子模块的步长均为1，卷积子模块的输入通道数为对应卷积模块的输入通道数，最大值池化层的卷积核为2*2；
所述基础识别网络的全连接层，输出通道数为512；
所述防伪分支的全连接层中，输出通道分别为16和2。
输入的指静脉图像如图3所示。
本发明使用的指静脉图像，使用红外光在手指一侧进行照射，并通过红外摄像头在另一侧成像，得到初步采集后的指静脉图像。然后需要对指静脉图像 进行预处理工作，这是因为在静脉采集装置上放置手指存在一定的灵活性，指静脉图像的偏移旋转引起静脉纹路的改变情况亦会存在。另外，采集的指静脉图像中含有大部分的背景信息，这些与指静脉无关的信息对后续识别会产生一定程度的干扰和影响，在预处理阶段需要进行手指图像的ROI截取以及旋转矫正，以尽量减少手指偏移和背景噪声带来的影响。
具体地说，对指静脉图像进行预处理如图4所示；具体步骤如下：
(1)在指静脉图像提取手指上下边缘；
因为背景和手指区域的边缘灰度差别较大，加上红外光源在边缘折射的光一般较透射光较强，手指边缘一般会出现明显的亮线，这种亮线大大加大了边缘检测的准确性。通过观察静脉图像可知，手指边缘的分界特征主要体现为上下边缘特征，这里采用普通的一阶微分算子即可，而因为手指左右方向即沿手指轴向方向并不需要考虑，因此这里仅使用垂直边缘模板即可。上边缘的特征是上方灰度值比下方灰度值大，而下边缘的特征则相反，下方灰度值比上方灰度值大，因此可以使用图5(a)所示的算子作为上边缘提取算子，图5(b)所示的算子作为下边缘提取算子进行提取。
(2)提取中线进行旋转矫正：
手指在放置时可能会出现水平旋转偏置，会对提取出的静脉纹路噪声较大的变化。而手指正常的姿势应是手指轴向在图像上接近水平，通过这一约束，可以通过上步骤提取的手指上下边缘计算出垂直的中点集合，然后通过最小二乘法拟合手指中线，从而求出手指与水平方向的倾斜角。然后通对图像旋转到相应角度，将手指矫正到水平方向。因为手指静脉主要集中在手指中心区域，通过计算旋转后手指的上下边缘的最低点和最高点，确定在垂直方向的截取轮廓线。
(3)指关节定位
由于用户将手指放入采集装置时，放入的位置存在一定的随机性，会导致手指在前后方向会存在变差，因此需要依靠稳定的参考信息才能截取到稳定的ROI。指静脉的采集一般采集的是食指，中指和无名指的图像，而这些手指都具有两个指骨间关节，这些指骨关节之间因存在关节液而存在较少的组织，其对红外光的吸收率较低，使得指关节区域的平均亮度大于其他手指区域，利用这一特征，可以较稳定地定位出用户在轴向的截取区域。根据这一特征，采用活 动窗口求和的方法来获取原始图像手指轴向方的亮度统计曲线趋势。活动窗口采用与原图一致的高度，和原图宽度1/20的宽度逐列滑动，计算窗口内的像素和。为了获得更加稳定的趋势序列，得到更加准确的指关节位置，这里先对数据进行高斯平滑处理，以排除异常值的干扰。在图中可以看出得到两个明显的波峰，分别对应手指的两个指间关节。这里为了算法更加鲁棒，本文只使用了最大峰值处的位置信息，向左延伸到起始点距离的1/3作为左端点，向右延伸图像长度的2/3，根据此区域进行水平方向截取，得到最终的ROI作为预处理后的指静脉数据。
在所述识别模式下，采用如下两种方式之一：
一、先判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若存在余弦距离大于距离阈值s 2时，则输出通过结果；否则输出拒绝结果；
二、先将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若不存在余弦距离大于距离阈值s 2时，则输出拒绝结果；否则继续判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则输出通过结果。
所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型，是指：
训练样本分为两个部分，分别为识别注册样本集合和防伪样本集合；其中，识别样本集合为R＝{r 1,r 2,...,r n}，防伪样本集合为S＝{s 1,s 2,...,s n}；防伪样本集合中的伪造样本，根据识别样本中的类别进行伪造生成，即s i＝f(r i)，0≤i＜n；因此其类别与识别样本中的类别具有一一对应关系。图6展示了伪造静脉和真实静脉的对比。
伪造生成方法为，选取识别样本集合其中一个类别中的样本，将该样本使用打印机进行打印到两张激光打印胶片上，然后将两张胶片堆叠在一起并对齐。在对齐好的两张胶片中放置一张高质量的白纸，制成伪造模型；然后将伪造模型放入静脉采集模型中采集，最终得到了伪造样本。
以遍历识别样本集合和防伪样本集合为迭代单位，交替对指静脉识别防伪任务卷积神经网络模型中的基础识别网络和防伪分支进行训练，如图7所示；在训练过程中，每次只有基础识别网络和防伪分支的其中一项参与训练，另一项的权重固定；降低任务间样本互相造成的性能影响。
在基础识别网络和防伪分支的训练中，均使用中心损失作为损失函数，其中中心损失为：
Figure PCTCN2020123182-appb-000002
其中，N表示样本数目，x代表网络输出的识别任务特征向量，c表示该类别的中心；
训练评价指标为：
GEER ω＝ω·SEER+(1-ω)·EER′
其中，ω表示基础识别网络和防伪分支的比重，SEER表示防伪分支的等误率，EER表示基础识别网络的等误率。
为评估本发明方法的性能，将本发明方法与现有的防伪方法进行对比。然后为验证算法的鲁棒性和泛化能力，在IDIAP防伪数据库和SCUT数据库上进行相关实验。为验证额外识别任务对防伪任务的影响，单独在防伪任务下训练相同结构的网络来进行对比验证。其结果如下表1所示。
表1
Figure PCTCN2020123182-appb-000003
Figure PCTCN2020123182-appb-000004
防伪任务作为一个二分类任务，其难度比识别任务相对较简单，因此深度网络可以相对有效地处理两种任务的结合，识别任务的加入对防伪任务的影响较少，提出的算法在识别和防伪结合任务上有其独有的优势。
同时，为评估识别与防伪结合网络在识别任务上的性能，在多个公开的指静脉数据库上进行实验，包括了IDIAP，USM，SDUMLA，MMCBNU和自建的SCUT数据集。同时为验证多任务训练的能力，选择与防伪任务类似的策略，将任务分为单独识别任务和识别防伪结合任务。最终的实验结果如下表2所示。
表2
Figure PCTCN2020123182-appb-000005
随着识别和防伪任务使用统一的模型进行特征处理，使用识别与防伪统一指标可以更为有效地评价系统的性能。为此，将识别任务的等误率EER和防伪任务的HTER指标使用权重进行组合，使用一种简化的权重组合方式GEER指标作为防伪与识别结合评价指标，其中ω表示防伪任务等误率的占比，其指标如下 表3所示。
表3
Figure PCTCN2020123182-appb-000006
另外，在时间消耗方面对提出的静脉识别防伪一体化模型进行评估。算法的部署平台在JetsonTK1开发板上进行，其算法的实现主要基于C++语言进行复现，深度模型的框架使用平台上的Tensorflow框架，最终的时间消耗如下表4所示。
表4
Figure PCTCN2020123182-appb-000007
提出的算法模型在100次前向运算的平均值为13.11，在部署算法的应用系统上，其实时性得到有效保障。
从防伪任务的角度，同时与多种防伪方法进行对比，证明了防伪任务受到识别任务的影响相对较少，其提出的算法模型能够获得较好的防伪性能。而在识别任务方面，本发明与多种传统方法与深度方法进行对比，在防伪任务加入后，其性能结果仍有较大竞争力。另外，还通过简化的防伪和识别任务指标，对整体的识别防伪一体化系统进行性能评估，并最终通过时耗实验来证明了算法具有较高的实时性。
实施例二
为实现实施例一所述的指静脉识别与防伪一体化方法，本实施例提供一种指静脉识别与防伪一体化装置，包括：
预处理模块，用于对指静脉图像进行预处理，得到预处理后的指静脉数据；
特征提取模块，用于将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
注册模块，用于实现注册模式，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
识别模块，用于实现识别模式，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
实施例三
本实施例一种存储介质，其特征在于，其中所述存储介质存储有计算机程序，所述计算机程序当被处理器执行时使所述处理器执行实施例一所述的指静脉识别与防伪一体化方法。
实施例四
本实施例一种计算设备，包括处理器以及用于存储处理器可执行程序的存储器，其特征在于，所述处理器执行存储器存储的程序时，实现实施例一所述的指静脉识别与防伪一体化方法。
上述实施例为本发明较佳的实施方式，但本发明的实施方式并不受上述实施例的限制，其他的任何未背离本发明的精神实质与原理下所作的改变、修饰、替代、组合、简化，均应为等效的置换方式，都包含在本发明的保护范围之内。

Claims (10)

  1. 一种指静脉识别与防伪一体化方法，其特征在于：包括：
    获取待识别的指静脉图像；对指静脉图像进行预处理，得到预处理后的指静脉数据；
    将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
    在注册模式下，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
    在识别模式下，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
  2. 根据权利要求1所述的指静脉识别与防伪一体化方法，其特征在于：所述指静脉识别防伪任务卷积神经网络模型包括基础识别网络和防伪分支；所述基础识别网络包括依次连接的两个卷积网络、三个卷积模块和全连接层；所述防伪分支包括依次连接的一个卷积模块和两个全连接层；防伪分支的卷积模块前端插入于基础识别网络的第一个卷积模块之后，从而构建出单输入多输出的指静脉识别防伪任务卷积神经网络模型。
  3. 根据权利要求2所述的指静脉识别与防伪一体化方法，其特征在于：所述基础识别网络的三个卷积模块均包括依次连接的两个卷积子模块和最大值池化层。
  4. 根据权利要求3所述的指静脉识别与防伪一体化方法，其特征在于：所述基础识别网络的两个卷积网络中，卷积核大小均为3*3，通道数均为64，步长分别为2和1；
    所述基础识别网络的三个卷积模块中，输入通道数分别为64、128和256；三个卷积模块中，前一卷积子模块的卷积核为3*3，后一卷积子模块的卷积核为1*1，卷积子模块的步长均为1，卷积子模块的输入通道数为对应卷积模块的输 入通道数，最大值池化层的卷积核为2*2；
    所述基础识别网络的全连接层，输出通道数为512；
    所述防伪分支的全连接层中，输出通道分别为16和2。
  5. 根据权利要求2所述的指静脉识别与防伪一体化方法，其特征在于：所述对指静脉图像进行预处理，是指：
    在指静脉图像提取手指上下边缘；提取手指上下边缘的垂直中点集合，通过最小二乘法拟合手指中线，从而求出手指与水平方向的倾斜角；然后对指静脉图像旋转，将手指矫正到水平方向；
    采用活动窗口求和的方法来获取原始指静脉图像手指轴向方的亮度统计曲线趋势；亮度统计曲线趋势的两个波峰设定为手指的两个指间关节；在两个指间关节之间截取ROI作为预处理后的指静脉数据。
  6. 根据权利要求2所述的指静脉识别与防伪一体化方法，其特征在于：所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型，是指：
    训练样本包括识别注册样本集合和防伪样本集合；识别样本集合为R＝{r 1,r 2,...,r n}，防伪样本集合为S＝{s 1,s 2,...,s n}；其中，防伪样本集合中的伪造样本，根据识别样本中的类别进行伪造生成，即s i＝f(r i)，0≤i＜n；
    以遍历识别样本集合和防伪样本集合为迭代单位，交替对指静脉识别防伪任务卷积神经网络模型中的基础识别网络和防伪分支进行训练；在训练过程中，每次只有基础识别网络和防伪分支的其中一项参与训练，另一项的权重固定；
    在基础识别网络和防伪分支的训练中，均使用中心损失作为损失函数，其中中心损失为：
    Figure PCTCN2020123182-appb-100001
    其中，N表示样本数目，x代表网络输出的识别任务特征向量，c表示该类别的中心；
    训练评价指标为：
    GEER ω＝ω·SEER+(1-ω)·EER′
    其中，ω表示基础识别网络和防伪分支的比重，SEER表示防伪分支的等误 率，EER表示基础识别网络的等误率。
  7. 根据权利要求1所述的指静脉识别与防伪一体化方法，其特征在于：在所述识别模式下，采用如下两种方式之一：
    一、先判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若存在余弦距离大于距离阈值s 2时，则输出通过结果；否则输出拒绝结果；
    二、先将指静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离进行逐一比较：若不存在余弦距离大于距离阈值s 2时，则输出拒绝结果；否则继续判断指静脉数据的防伪任务分类概率p与概率阈值s 1之间大小：若指静脉数据的防伪任务分类概率p＜阈值s 1，则判定指静脉图像为伪造样本，并输出拒绝结果；否则输出通过结果。
  8. 一种指静脉识别与防伪一体化装置，其特征在于，包括：
    预处理模块，用于对指静脉图像进行预处理，得到预处理后的指静脉数据；
    特征提取模块，用于将所述预处理后的指静脉数据输入至指静脉识别防伪任务卷积神经网络模型，通过所述指静脉识别防伪任务卷积神经网络模型对所述指静脉图像进行识别和防伪处理，得到防伪任务分类概率p和识别任务特征向量v；其中，所述指静脉识别防伪任务卷积神经网络模型为对初始指静脉识别防伪任务卷积神经网络模型进行训练处理得到的模型；
    注册模块，用于实现注册模式，当指静脉数据的防伪任务分类概率p≤概率阈值s 1，则将静脉数据的识别任务特征向量v输出并保存为注册样本识别任务特征向量；
    识别模块，用于实现识别模式，通过对静脉数据的防伪任务分类概率p与概率阈值s 1进行比较，以及对静脉数据的识别任务特征向量v与各个注册样本识别任务特征向量的余弦距离与距离阈值s 2进行比较，输出判定结果。
  9. 一种存储介质，其特征在于，其中所述存储介质存储有计算机程序，所述计算机程序当被处理器执行时使所述处理器执行权利要求1-7中任一项所述的指静脉识别与防伪一体化方法。
  10. 一种计算设备，包括处理器以及用于存储处理器可执行程序的存储器， 其特征在于，所述处理器执行存储器存储的程序时，实现权利要求1-7中任一项所述的指静脉识别与防伪一体化方法。
PCT/CN2020/123182 2020-06-05 2020-10-23 指静脉识别与防伪一体化方法、装置、存储介质和设备 WO2021243926A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2020432845A AU2020432845B2 (en) 2020-06-05 2020-10-23 An integrated method for finger vein recognition and anti-spoofing, and device, storage medium and equipment therefor

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010505303.4 2020-06-05
CN202010505303.4A CN111914616B (zh) 2020-06-05 2020-06-05 指静脉识别与防伪一体化方法、装置、存储介质和设备

Publications (1)

Publication Number Publication Date
WO2021243926A1 true WO2021243926A1 (zh) 2021-12-09

Family

ID=73238045

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/123182 WO2021243926A1 (zh) 2020-06-05 2020-10-23 指静脉识别与防伪一体化方法、装置、存储介质和设备

Country Status (3)

Country Link
CN (1) CN111914616B (zh)
AU (1) AU2020432845B2 (zh)
WO (1) WO2021243926A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114913610A (zh) * 2022-06-15 2022-08-16 南京邮电大学 一种基于指纹和指静脉的多模态识别方法
CN116363712A (zh) * 2023-03-21 2023-06-30 中国矿业大学 一种基于模态信息度评估策略的掌纹掌静脉识别方法
CN117218686A (zh) * 2023-10-20 2023-12-12 广州脉泽科技有限公司 一种开放场景下的掌静脉roi提取方法及系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112766142B (zh) * 2021-01-15 2022-12-20 天津大学 足底压力图像处理方法、识别方法及步态分析系统
CN114821825B (zh) * 2022-06-30 2022-12-06 广州中平智能科技有限公司 一种多粒度人脸伪造检测方法、系统、设备和介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110304720A1 (en) * 2010-06-10 2011-12-15 The Hong Kong Polytechnic University Method and apparatus for personal identification using finger imaging
CN106295555A (zh) * 2016-08-08 2017-01-04 深圳芯启航科技有限公司 一种活体指纹图像的检测方法
CN109522798A (zh) * 2018-10-16 2019-03-26 平安科技（深圳）有限公司 基于活体识别的视频防伪方法、系统、装置及可存储介质
CN110717372A (zh) * 2019-08-13 2020-01-21 平安科技（深圳）有限公司 基于指静脉识别的身份验证方法和装置
CN111178221A (zh) * 2019-12-24 2020-05-19 珠海格力电器股份有限公司 身份识别方法及装置

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110390282A (zh) * 2019-07-12 2019-10-29 西安格威西联科技有限公司 一种基于余弦中心损失的指静脉识别方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110304720A1 (en) * 2010-06-10 2011-12-15 The Hong Kong Polytechnic University Method and apparatus for personal identification using finger imaging
CN106295555A (zh) * 2016-08-08 2017-01-04 深圳芯启航科技有限公司 一种活体指纹图像的检测方法
CN109522798A (zh) * 2018-10-16 2019-03-26 平安科技（深圳）有限公司 基于活体识别的视频防伪方法、系统、装置及可存储介质
CN110717372A (zh) * 2019-08-13 2020-01-21 平安科技（深圳）有限公司 基于指静脉识别的身份验证方法和装置
CN111178221A (zh) * 2019-12-24 2020-05-19 珠海格力电器股份有限公司 身份识别方法及装置

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114913610A (zh) * 2022-06-15 2022-08-16 南京邮电大学 一种基于指纹和指静脉的多模态识别方法
CN116363712A (zh) * 2023-03-21 2023-06-30 中国矿业大学 一种基于模态信息度评估策略的掌纹掌静脉识别方法
CN116363712B (zh) * 2023-03-21 2023-10-31 中国矿业大学 一种基于模态信息度评估策略的掌纹掌静脉识别方法
CN117218686A (zh) * 2023-10-20 2023-12-12 广州脉泽科技有限公司 一种开放场景下的掌静脉roi提取方法及系统
CN117218686B (zh) * 2023-10-20 2024-03-29 广州脉泽科技有限公司 一种开放场景下的掌静脉roi提取方法及系统

Also Published As

Publication number Publication date
AU2020432845B2 (en) 2023-03-16
CN111914616A (zh) 2020-11-10
CN111914616B (zh) 2024-04-16
AU2020432845A1 (en) 2021-12-23

Similar Documents

Publication Publication Date Title
WO2021243926A1 (zh) 指静脉识别与防伪一体化方法、装置、存储介质和设备
WO2017059591A1 (zh) 手指静脉识别方法及装置
CN104239769B (zh) 基于手指静脉特征的身份识别方法及系统
CN102542281B (zh) 非接触式生物特征识别方法和系统
CN100492400C (zh) 手指静脉特征提取与匹配识别方法
CN101055618A (zh) 基于方向特征的掌纹识别方法
CN101030244A (zh) 基于人体生理图像中排序测度特征的自动身份识别方法
Zhang et al. Graph fusion for finger multimodal biometrics
CN109544523B (zh) 基于多属性人脸比对的人脸图像质量评价方法及装置
CN105760841A (zh) 一种身份识别方法及系统
CN1912889A (zh) 基于局部三角结构特征集的形变指纹识别方法
CN1421815A (zh) 基于知识的指纹图像增强方法
CN104091145A (zh) 人体掌脉特征图像采集方法
CN113469143A (zh) 一种基于神经网络学习的手指静脉图像识别方法
Kisku et al. Multispectral palm image fusion for person authentication using ant colony optimization
Qin et al. Finger-vein image quality evaluation based on the representation of grayscale and binary image
Ting et al. A review of finger vein recognition system
Prashanth et al. Off-line signature verification based on angular features
Latha et al. A robust person authentication system based on score level fusion of left and right irises and retinal features
CN1457015A (zh) 基于人脸识别和手形识别的双模态生物认证系统
Akulwar et al. Secured multi modal biometric system: a review
Khoirunnisaa et al. The biometrics system based on iris image processing: a review
Gopinath et al. Exploration of finger vein recognition systems
Sujatha et al. Multimodal biometric authentication algorithm at score level fusion using hybrid optimization
Radouane et al. Fusion of Gabor filter and steerable pyramid to improve iris recognition system

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2020432845

Country of ref document: AU

Date of ref document: 20201023

Kind code of ref document: A

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

Ref document number: 20938622

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 28.03.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20938622

Country of ref document: EP

Kind code of ref document: A1