WO2023029274A1 - 基于多方合作的指静脉识别方法、***和存储介质 - Google Patents
基于多方合作的指静脉识别方法、***和存储介质 Download PDFInfo
- Publication number
- WO2023029274A1 WO2023029274A1 PCT/CN2021/136425 CN2021136425W WO2023029274A1 WO 2023029274 A1 WO2023029274 A1 WO 2023029274A1 CN 2021136425 W CN2021136425 W CN 2021136425W WO 2023029274 A1 WO2023029274 A1 WO 2023029274A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- finger vein
- model
- training
- module
- algorithm
- Prior art date
Links
- 210000003462 vein Anatomy 0.000 title claims abstract description 82
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000003860 storage Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 57
- 230000007246 mechanism Effects 0.000 claims abstract description 38
- 238000001514 detection method Methods 0.000 claims description 12
- 238000007781 pre-processing Methods 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 9
- 238000013527 convolutional neural network Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 238000004140 cleaning Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 238000002372 labelling Methods 0.000 claims description 4
- 238000013480 data collection Methods 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 description 12
- 238000011176 pooling Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 6
- 230000004913 activation Effects 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000000712 assembly Effects 0.000 description 1
- 238000000429 assembly Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 239000010977 jade Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Images
Classifications
-
- 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
- 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
Definitions
- the invention belongs to biometric identification technology, relates to the field of artificial intelligence based on machine vision, in particular to a finger vein identification method, system and storage medium based on multi-party cooperation.
- Finger vein recognition technology achieves the purpose of identification by analyzing the texture of the veins in the fingers. Compared with identification technology, such as face recognition, fingerprint recognition, etc., finger vein recognition technology has higher security, because the vein pattern in the finger belongs to the internal information of the human body, which cannot be copied or forged. Therefore, as one of the currently recognized biometric technologies with a high safety factor, finger vein recognition technology has been widely used in banking systems, insurance industries and other financial fields.
- the recognition algorithm is not perfect. As the most important link in the finger vein recognition technology, the finger vein recognition algorithm directly affects the accuracy and efficiency of recognition. Algorithms currently on the market often directly use networks such as ResNet, VGG, MobileNet, and ShuffleNet to recognize finger vein patterns. However, deep neural networks such as ResNet and VGG have excellent recognition performance, but the models are complex and the recognition time is long, which is not conducive to It is directly applied to terminal devices; while the recognition effect of lightweight network structures such as MobileNet and ShuffleNet needs to be improved urgently. See Figure 1. After such operations as convolution, normalization, and pooling, such lightweight networks can obtain representative indicators. The depth semantic information of the vein pattern is extracted.
- the data size of the training set is small.
- Current neural network-based algorithms often need to use large-scale training data sets.
- the larger the training data set the better the performance of the trained model under the same conditions.
- some small and medium-sized enterprises are unable to collect large-scale data sets due to resource constraints, and can only use small-scale data sets to train the model, but the training strategy of using small-scale data set training often leads to poor generalization ability of the model , the robustness is not good, and it can only adapt to the scene very similar to the training data set.
- the algorithm does not match the data format of the application terminal.
- the data format includes the aspect ratio of the image saved during the collection of the finger vein image, the image format type (.jpg,, bmp, .tif, etc.), the ratio of the finger vein target area to the total image area, and so on. Due to the diversification of terminals used in the finger vein recognition products currently on the market when collecting finger vein lines, the collected data formats have different attributes, and the current algorithm formed by training a single data set cannot adapt to the different data formats of multiple terminals. Higher accuracy is only applicable to the data format similar to the training data set, so it is necessary to eliminate the recognition error caused by the different data formats of the algorithm and the multi-application terminal.
- the current finger vein recognition technology still has many deficiencies in the process of market-oriented application and needs to be further improved.
- the purpose of the present invention is to provide a finger vein recognition method, system and storage medium based on multi-party cooperation, which can solve the above problems.
- This patent proposes a lightweight finger vein recognition algorithm based on the attention mechanism. Compared with the conventional recognition algorithm, it has The recognition algorithm of the attention mechanism can pay more attention to the network structure of finger veins and suppress other interference factors; 2 Aiming at the problems of small training data set and mismatch between algorithm and multi-terminal data format, this patent proposes a multi-user collaborative training framework.
- the conventional training method is often a single-user mode, that is, a single user, a single training data set, and a single training data format, and the proposal of the multi-user collaborative training framework can train multiple users together without leaking corporate secrets.
- a finger vein recognition method based on multi-party cooperation comprising:
- S1. Data collection collect finger vein images, and perform data labeling and cleaning on the collected images, so as to standardize the training data set;
- Model training first establish a training model that introduces the attention mechanism, and then based on the multi-user collaborative training framework, multiple partners jointly train the optimized global model without data exchange;
- Model testing the global model optimized in step S2 is used as the detection model, and the test data set is brought into the detection model for detection.
- step S2 the initially established training model adopts a convolutional neural network structure, and an attention mechanism is added after each part to form a network structure with an attention mechanism.
- the convolutional neural network structure adopts a lightweight network MobileNet or ShuffleNet algorithm model
- the attention mechanism adopts a CBAM attention mechanism including a spatial attention mechanism module and a channel attention mechanism module.
- obtaining the optimized global model in step S2 includes the following training steps:
- the partner updates the local model and uses local data for training to obtain N local models;
- Each partner performs homomorphic encryption on the local model parameters, and sends the encrypted model parameters to the cooperation center;
- the cooperation center aggregates N model parameters to obtain a global model, and returns the global model parameters to each partner;
- Each partner decrypts the returned model parameters, and repeats S222-S224 until the model converges or reaches the preset number of training rounds, so as to obtain an optimized global model.
- step S223 the parameters of the local model are encrypted through a homomorphic encryption algorithm.
- the test data set includes the finger vein model atlas of legal users and the finger vein model graph of illegal users.
- the present invention also provides a finger vein identification system, the system includes a finger vein image acquisition module, an image preprocessing module, an identity authentication module, a storage module and a result output module, the finger vein image acquisition module adopts a finger vein acquisition instrument, It is telecommunications connected with the image preprocessing module and the storage module; the image preprocessing module preprocesses the finger vein image pre-stored in the storage module; The detection model is to judge the legitimacy of the preprocessed image extraction features, and judge whether it is a legitimate user by comparing with the legal finger vein image backed up by the storage module; A vein recognition method and a newly entered finger vein image of a user to be detected; the result output module includes a legal pass unit and an illegal alarm unit.
- the present invention also provides a computer-readable storage medium, on which computer instructions are stored, and the steps of the aforementioned finger vein recognition method are executed when the computer instructions are executed.
- the beneficial effect of the present invention is that: through this application, in the recognition algorithm, a lightweight finger vein recognition algorithm based on the attention mechanism is introduced, and the attention mechanism can pay more attention to the network structure and suppress other interference factors, which While the algorithm has high recognition accuracy, it consumes low computing power, thereby improving the performance of the algorithm; using a multi-user collaborative training framework, it can not only not leak secrets, but also can be trained jointly through the cooperation center, which can improve the performance of the model. robustness.
- Fig. 1 is a schematic diagram of a conventional recognition algorithm model
- Fig. 2 is a schematic diagram of the recognition algorithm model based on attention of the present invention
- Figure 3 is a schematic diagram of the improved algorithm
- Figure 4 is a schematic diagram of the principle of the CBAM attention mechanism
- Figure 5 is a schematic diagram of a multi-user collaborative training framework
- FIG. 6 is a flow chart of the overall method of the finger vein recognition method
- Fig. 7 is a schematic diagram of the recognition system.
- system means for distinguishing different components, elements, parts, parts or assemblies of different levels.
- the words may be replaced by other expressions if other words can achieve the same purpose.
- a finger vein recognition method based on multi-party cooperation see Figure 6, the method includes:
- S1. Data collection collect finger vein images, and perform data labeling and cleaning on the collected images, so as to standardize the training data set.
- finger vein images should be collected as much as possible, and transfer learning can also be applied to reduce the sample size.
- Image cleaning includes ROI interception, image enhancement, normalization, etc. It should also include feature extraction for labeling, such as binarization processing, edge interception, rotation correction, joint positioning, maximum principal curvature, etc.
- Model training first establish a training model that introduces an attention mechanism, and then based on the multi-user collaborative training framework, multiple partners jointly train an optimized global model without data exchange.
- step S3 Model testing: the global model optimized in step S2 is used as the detection model, and the test data set is brought into the detection model for detection.
- the test data set includes the finger vein model atlas of legal users and the finger vein model graph of illegal users.
- step S2 the initially established training model adopts a convolutional neural network structure, and an attention mechanism is added after each part, as shown in Fig. 2 and Fig. 3 , to form a network structure with an attention mechanism.
- the convolutional neural network structure adopts a lightweight network MobileNet or ShuffleNet algorithm model
- the attention mechanism adopts a CBAM attention mechanism including a spatial attention mechanism module and a channel attention mechanism module.
- An algorithm model based on the attention mechanism is designed, and it is used in convolution, normalization, pooling, etc. During the operation, the attention mechanism is added, and the deep semantic information representing the finger vein pattern is extracted. The attention mechanism model can effectively suppress the noise in the finger vein image and pay attention to the texture of the finger vein.
- the CBAM attention mechanism can focus on finger veins from two aspects: spatial features (spatial attention mechanism module focuses on feature map, pixel-level finger vein texture) and channel features (channel attention mechanism module focuses on more expressive feature channels) Textures can improve the recognition ability of the network, so the CBAM attention mechanism can be used in this scheme.
- the network structure of the CBAM attention mechanism is shown in Figure 4.
- F in enters the feature channel attention mechanism module channel attention block
- MLP weights the channel features of the two feature maps to obtain the weighted maximum pooling feature and average pooling feature, and then adds the two features for data activation, and finally passes the channel attention feature matrix obtained after activation (channel attention feature) is multiplied by the input feature to obtain the channel-refined feature (channel-refined feature).
- channel-refined features are first concatenated directly after max pooling (MaxPool) and average pooling (AvgPool), and then convolution and activation are performed to obtain the spatial attention feature matrix (Spatial attention feature), and finally the spatial attention feature matrix is multiplied by the channel fine-tuning feature to obtain the output F out .
- MaxPool max pooling
- AvgPool average pooling
- F out the spatial attention feature matrix
- F in represents the input feature
- F out represents the output feature
- M S ( ⁇ ) represents the mapping function of the spatial attention module
- M c ( ⁇ ) represents the mapping function of the channel attention module
- the specific expressions are formula (2), formula (3):
- ⁇ ( ⁇ ) represents the activation function
- MLP( ⁇ ) represents the multi-layer perceptron
- f 7 ⁇ 7 ( ⁇ ) represents the convolution operation with a convolution kernel of 7 ⁇ 7
- Cat( ⁇ ) represents the Cascading (splicing)
- AvgPool( ) means average pooling
- MaxPool( ) means maximum pooling.
- step S2 For the optimized global model, to solve the two problems of small training data set size and mismatch between algorithm and multi-terminal data format, this solution designs a multi-user collaborative training framework, as shown in Figure 5, where the cooperation center performs model aggregation.
- the partner encrypts and decrypts the model, and trains the local model.
- obtaining the optimized global model in step S2 includes the following training steps.
- the partner updates the local model and uses the local data for training to obtain N local models.
- Each partner performs homomorphic encryption on the local model parameters, and sends the encrypted model parameters to the cooperation center.
- the cooperation center aggregates the N model parameters to obtain a global model, and returns the global model parameters to each partner.
- Each partner decrypts the returned model parameters, and repeats S222-S224 until the model converges or reaches the preset number of training rounds, so as to obtain an optimized global model.
- step S223 the parameters of the local model are encrypted by a homomorphic encryption algorithm.
- the homomorphic encryption algorithm ensures that the decryption result of the partner's ciphertext is equivalent to the partner's plaintext. While ensuring the security of the data content, joint training is performed to generate a jointly optimized model. Commonly used homomorphic encryption algorithms include RAS algorithm, Paillier algorithm, etc. RAS algorithm can be used in this patent.
- the algorithm process includes:
- the ciphertext is c
- This training method guarantees data privacy, and can be jointly trained without exchanging data, and the model can be jointly optimized without the data of each partner being leaked, and the global model can be collaboratively trained using the data of each partner to improve Model effect, resist external attacks by encrypting model parameters during transmission.
- the participation of multiple partners represents the entry of multiple terminal devices, which indirectly solves the problem of multi-terminal data mismatch.
- this scheme proposes a multi-user collaborative training framework.
- the conventional training method is often a single-user mode, that is, a single user, a single training data set, and a single training data format, and the proposal of the multi-user collaborative training framework can train multiple users together without leaking corporate secrets.
- a finger vein identification system see FIG. 7, the system includes a finger vein image acquisition module, an image preprocessing module, an identity authentication module, a storage module and a result output module.
- the finger vein image acquisition module adopts a finger vein acquisition instrument, which is connected with the image preprocessing module and the storage module by telecommunication.
- the image preprocessing module preprocesses the finger vein images pre-stored in the storage module.
- the identity authentication module uses the detection model in the aforementioned finger vein recognition method to judge the legitimacy of the preprocessed image extraction features, and judges whether it is a legitimate user by comparing with the legal finger vein image backed up by the storage module.
- the storage module stores the legitimate user's finger vein image and user information, the finger vein identification method and the newly entered finger vein image of the user to be detected.
- the result output module includes a legal pass unit and an illegal alarm unit.
- Legal pass units such as access control, illegal alarm units such as buzzers, warning lights and or voice prompts.
- the present invention also provides a storage medium on which computer instructions are stored, and the steps of the aforementioned finger vein recognition method are executed when the computer instructions are run.
- a storage medium on which computer instructions are stored, and the steps of the aforementioned finger vein recognition method are executed when the computer instructions are run.
- Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Flash memory or other memory technology, Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cartridge, tape magnetic disk storage or other magnetic storage device or any other non-transmission medium that can be used to store information that can be accessed by a computing device.
- computer-readable media excludes transitory computer-readable media, such as modulated data signals and carrier waves.
- the computer program codes required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python etc., conventional procedural programming languages such as C language, VisualBasic, Fortran2003, Perl, COBOL2002, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy, or other programming languages.
- the program code may run entirely on the user's computer, or as a stand-alone software package, or run partly on the user's computer and partly on a remote computer, or entirely on the remote computer or processing device.
- the remote computer can be connected to the user computer through any form of network, such as a local area network (LAN) or wide area network (WAN), or to an external computer (such as through the Internet), or in a cloud computing environment, or as a service Use software as a service (SaaS).
- LAN local area network
- WAN wide area network
- SaaS service Use software as a service
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Collating Specific Patterns (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
本发明提供了一种基于多方合作的指静脉识别方法、***和存储介质,属于生物识别技术领域,在识别算法中,引入了基于注意力机制的轻量化指静脉识别算法,注意力机制能够更加关注网络结构抑制其他干扰因素,这使算法在具有较高的识别准确性的同时,消耗的算力低,以此改善算法的性能;采用多用于协同训练框架,既能***密又能通过合作中心协同共同训练,能够提高模型的鲁棒性。
Description
本发明属于生物识别技术,涉及基于机器视觉的人工智能领域,具体涉及一种基于多方合作的指静脉识别方法、***和存储介质。
本部分的陈述仅仅是提供了与本发明相关的背景技术信息,不必然构成在先技术或先有技术。
指静脉识别技术通过分析手指中静脉的纹路达到身份识别的目的。相比于身份识别技术,比如人脸识别,指纹识别等,指静脉识别技术具有较高的安全性,因为手指中静脉的纹路属于人体内部信息,无法复制,不可伪造。因此,作为目前公认的安全系数较高的生物识别技术之一,指静脉识别技术在银行***、保险行业等金融领域有着广泛的应用。
然而,在当前生产实践的过程中,指静脉识别技术无法得到大面积的推广,还存在着识别算法的不完善、训练数据集规模小以及算法与多终端数据格式不匹配等问题,具体如下。
1、识别算法不完善。指静脉识别算法作为指静脉识别技术中最重要的环节,直接影响了识别的准确性以及识别效率。当前市场上的算法往往直接使用ResNet、VGG、MobileNet、ShuffleNet等网络对指静脉纹路进行识别,然而ResNet、VGG等深度神经网络虽然具有优秀的识别性能,但是模型复杂,识别时间较长,不利于直接应用于终端设备;而MobileNet、ShuffleNet等轻量化网络结构的识别效果又亟待提高,参见图1,这类轻量网络在经过卷积、归一化、池化等操作之后,能够获得代表指静脉纹路的 深度语义信息被提取出来。然而,受到噪声、采集设备不理想、指纹干扰等原因,影响了训练数据集的采集,获得的数据不干净,间接的干扰了模型的训练。因此需要一种强有力的识别技术,既能兼顾模型的准确性又能兼顾模型的识别算力(模型复杂度,识别时间)。
2、训练集数据规模小。当前基于神经网络的算法往往需要使用大规模的训练数据集,一般来说,训练数据集的规模越大,同等条件下训练得到的模型的性能越好。然而部分中小企业受到资源的限制而无法采集大规模的数据集,同时也只能采用小规模数据集对模型进行训练,但是采用小规模数据集训练的训练策略往往导致模型的泛化能力较差,鲁棒性不好,只能适应跟训练数据集极为相似的场景。
3、算法与应用终端数据格式的不匹配。数据格式包括采集指静脉图像过程中保存的图像的长宽比,图像格式类型(.jpg,,bmp,.tif等),指静脉目标区域占图像总区域的比率等等。由于目前市场上使用的指静脉识别产品时在采集指静脉纹路时的终端是多样化的,采集的数据格式具有不同属性,而目前单一数据集训练形成的算法无法针对多终端不同的数据格式具有较高的准确性,仅仅只适用与训练数据集相似的数据格式,因此需要消除算法和多应用终端数据格式不一导致的识别误差。
综上所诉,现在的指静脉识别技术在面向市场应用化过程中还存在诸多不足,需进一步改善。
发明内容
为了克服现有技术的不足,本发明的目的在于提供一种基于多方合作的指静脉识别方法、***和存储介质,其能解决上述问题。
设计原理:①针对识别算法的不完善,即基于轻量化模型指静脉算法 的识别性能不佳,本专利提出了基于注意力机制的轻量化指静脉识别算法,较常规的识别算法而言,具有注意力机制的识别算法能够更加关注指静脉的网络结构,抑制其他干扰因素;②针对训练数据集较小以及算法与多终端数据格式不匹配的问题,本专利提出了多用户协同训练框架。常规的训练方式,往往是单用户模式,即单一用户、单一训练数据集、单一的训练数据格式,而多用户协同训练框架的提出,能够将多用户在保证企业机密***露的前提下共同训练,因此也聚合了不同数据集的属性,提高了模型的鲁棒性,而且在训练过程中多用户将对适应于自己数据格式特征的模型传递给合作中心共同训练模型,也避免了算法仅仅只适应于单一的数据格式的现象。
总体方案:为了解决上述问题,本申请的总体设计方案如下。
一种基于多方合作的指静脉识别方法,方法包括:
S1、数据采集:采集指静脉图像,并对采集的图像进行数据标注和清洗,以此规范化训练数据集;
S2、模型训练:首先建立引入注意力机制的训练模型,再基于多用户协同训练框架,多合作方之间在无数据交换的情况下共同训练得到优化的全局模型;
S3、模型测试:以步骤S2优化后的全局模型作为检测模型,将测试数据集带入检测模型进行检测。
进一步的,在步骤S2中,初始建立的训练模型采用卷积神经网络结构,且每一部分之后都增加一个注意力机制,形成具有注意力机制的网络结构。
进一步的,所述卷积神经网络结构采用轻量网络的MobileNet或 ShuffleNet算法模型,注意力机制采用包括空间注意力机制模块和通道注意力机制模块的CBAM注意力机制。
进一步的,步骤S2中优化的全局模型获得包括以下训练步骤:
S221、合作方初始化模型参数;
S222、合作方更新本地模型并使用本地数据进行训练,得到N个本地模型;
S223、各个合作方分别对本地模型参数进行同态加密,将加密后的模型参数发送给合作中心;
S224、合作中心将N个模型参数进行聚合得到全局模型,将全局模型参数返回给各个合作方;
S225、各个合作方对返回的模型参数解密,重复S222~S224直至模型收敛或达到预设训练轮数,以此获得优化的全局模型。
进一步的,在步骤S223中,本地模型的参数通过同态加密算法进行加密。
进一步的,在步骤S3中测试数据集包括合法用户的手指静脉模型图集和非法用户的手指静脉模型图。
本发明还提供了一种指静脉身份识别***,***包括指静脉图像采集模块、图像预处理模块、身份认证模块、存储模块和结果输出模块,所述指静脉图像采集模块采用指静脉采集仪,其与图像预处理模块和存储模块电信连接;所述图像预处理模块将预存在存储模块的指静脉图像进行预处理;所述身份认证模块采用前述任一项所述的指静脉识别方法中的检测模型,将预处理后的图像提取特征进行合法性判断,通过与存储模块备份的 合法指静脉图像比较判断是否为合法用户;所述存储模块存储有合法用户的指静脉图像和用户信息、指静脉识别方法以及新录入的待检测用户的指静脉图像;所述结果输出模块包括合法通过单元和非法报警单元。
本发明还提供了一种计算机可读存储介质,其上存储有计算机指令,所述计算机指令运行时执行前述指静脉识别方法的步骤。
相比现有技术,本发明的有益效果在于:通过本申请,在识别算法中,引入了基于注意力机制的轻量化指静脉识别算法,注意力机制能够更加关注网络结构抑制其他干扰因素,这使算法在具有较高的识别准确性的同时,消耗的算力低,以此改善算法的性能;采用多用户协同训练框架,既能***密又能通过合作中心协同共同训练,能够提高模型的鲁棒性。
图1为常规的识别算法模型示意图;
图2为本发明基于注意力的识别算法模型示意图;
图3为改进的算法示意图;
图4为CBAM注意力机制原理示意图;
图5为多用户协同训练框架示意图;
图6为指静脉识别方法的整体方法流程图;
图7为识别***的示意图。
为使本发明实施例的目的、技术方案和优点更加清楚,下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本发明的一部分实施例,而不是全部的实施 例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例,都属于本发明保护的范围。
应当理解,本说明书中所使用的“***”、“装置”、“单元”和/或“模组”是用于区分不同级别的不同组件、元件、部件、部分或装配的一种方法。然而,如果其他词语可实现相同的目的,则可通过其他表达来替换所述词语。
本说明书中使用了流程图用来说明根据本说明书的实施例的***所执行的操作。应当理解的是,前面或后面操作不一定按照顺序来精确地执行。相反,可以按照倒序或同时处理各个步骤。同时,也可以将其他操作添加到这些过程中,或从这些过程移除某一步或数步操作。
识别方法
一种基于多方合作的指静脉识别方法,参见图6,方法包括:
S1、数据采集:采集指静脉图像,并对采集的图像进行数据标注和清洗,以此规范化训练数据集。
此处指静脉图像应尽量多采集,也可以应用迁移学习以减少样本量。图像的清洗包括ROI截取、图像增强、归一化等,还应包括特征提取进行标注,如二值化处理、边缘截取、旋转矫正、关节定位、最大主曲率等等。
S2、模型训练:首先建立引入注意力机制的训练模型,再基于多用户协同训练框架,多合作方之间在无数据交换的情况下共同训练得到优化的全局模型。
S3、模型测试:以步骤S2优化后的全局模型作为检测模型,将测试数据集带入检测模型进行检测。在步骤S3中测试数据集包括合法用户的手指静脉模型图集和非法用户的手指静脉模型图。
其中,在步骤S2中,初始建立的训练模型采用卷积神经网络结构,且每一部分之后都增加一个注意力机制,参见图2和图3,形成具有注意力机制的网络结构。
所述卷积神经网络结构采用轻量网络的MobileNet或ShuffleNet算法模型,注意力机制采用包括空间注意力机制模块和通道注意力机制模块的CBAM注意力机制。
该部分是为了改善算法的性能,使算法在具有较高的识别准确性的同时,消耗的算力较低,设计了基于注意力机制的算法模型,在卷积、归一化、池化等操作时,增加注意力机制,获得代表指静脉纹路的深度语义信息被提取出来。该注意力机制模型,能够有效的抑制指静脉图像中的噪声,关注指静脉的纹路。
CBAM注意力机制能够从空间特征(空间注意力机制模块关注特征图中,像素级别的指静脉纹理)和通道特征(通道注意力机制模块关注更具表达能力的特征通道)两个方面关注指静脉纹路,能够提高网络的识别能力,因此在本方案中可采用CBAM注意力机制。
CBAM注意力机制的网络结构参见图4,F
in输入特征通道注意力机制模块(channel attention block)时,首先分别经过最大池化(MaxPool)和平均池化(AvgPool),然后通过多层感知机(MLP)对两种特征图的通道特征进行加权,获得加权后的最大池化特征和平均池化特征,随后两种特征相加,进行数据激活,最后通过激活后获得的通道注意力特征矩阵(channel attention feature)与输入特征相乘得到通道微调特征(channel-refined feature)。而在空间注意力模块中通道微调特征(channel-refined feature)首先经过最大池化(MaxPool)和平均池化(AvgPool)之后直接拼接,随后经过卷积 以及激活获得空间注意力特征矩阵(Spatial attention feature),最后空间注意力特征矩阵与通道微调特征相乘,获得输出F
out。具体而言:
M
s(x)=σ(f
7×7(Cat(AvgPool(x),MaxPool(x)))) (3)
式中,σ(·)表示激活函数,MLP(·)表示多层感知机,f
7×7(·)表示卷积核为7×7的卷积操作,Cat(·)表示特征之间的级联(拼接),AvgPool(·)表示平均池化,MaxPool(·)表示最大池化。
对于优化的全局模型,是针对训练数据集规模较小以及算法和多终端数据格式不匹配这两个问题,本方案设计了多用户协同训练框架,参见图5,其中,合作中心进行模型聚合。合作方对模型进行加密、解密,和训练本地模型。具体的,步骤S2中优化的全局模型获得包括以下训练步骤。
S221、合作方初始化模型参数。
S222、合作方更新本地模型并使用本地数据进行训练,得到N个本地模型。
S223、各个合作方分别对本地模型参数进行同态加密,将加密后的模型参数发送给合作中心。
S224、合作中心将N个模型参数进行聚合得到全局模型,将全局模型参数返回给各个合作方。
S225、各个合作方对返回的模型参数解密,重复S222~S224直至模型收敛或达到预设训练轮数,以此获得优化的全局模型。
在步骤S223中,本地模型的参数通过同态加密算法进行加密。同态加密算法保证了对合作方密文的解密结果等价于合作方明文,在保证数据内容安全的同时,进行联合训练,生成共同优化后的模型。常用的同态加密算法有RAS算法,Paillier算法等,在本专利中可采用RAS算法,算法过程包括:
a、产生公钥和私钥:
随意选择两个大的质数p和q,p不等于q,令N=pq,r=(p-1)(q-1);随机选择一个小于r的整数e(满足gcd(e,r)=1),求得e关于r的模反元素d(即e×d=1mod r),则有公钥(N,e),私钥(N,d);
b、明文m,加密为c:c=m
e modN;
c、密文为c,解密为m:m=c
d modN。
该种训练方式保证了数据隐私,可以在不交换数据的情况下共同训练,在各个合作方的数据***露的情况下,共同优化模型,利用各合作方的数据协作的训练出全局模型,提升模型效果,传输过程中通过对模型参数加密抵抗外界攻击。此外,多合作方的参与代表了多终端设备的进入,间接的解决了多终端数据不匹配的问题。
综上,针对识别算法的不完善,即基于轻量化模型指静脉算法的识别性能不佳,本方案提出了基于注意力机制的轻量化指静脉识别算法,较常规的识别算法而言,具有注意力机制的识别算法能够更加关注指静脉的网络结构,抑制其他干扰因素;同时,针对训练数据集较小以及算法与多终端数据格式不匹配的问题,本方案提出了多用户协同训练框架。常规的训 练方式,往往是单用户模式,即单一用户、单一训练数据集、单一的训练数据格式,而多用户协同训练框架的提出,能够将多用户在保证企业机密***露的前提下共同训练,因此也聚合了不同数据集的属性,提高了模型的鲁棒性,而且在训练过程中多用户将对适应于自己数据格式特征的模型传递给合作中心共同训练模型,也避免了算法仅仅只适应于单一的数据格式的现象。
识别***
一种指静脉身份识别***,参见图7,***包括指静脉图像采集模块、图像预处理模块、身份认证模块、存储模块和结果输出模块。
所述指静脉图像采集模块采用指静脉采集仪,其与图像预处理模块和存储模块电信连接。
所述图像预处理模块将预存在存储模块的指静脉图像进行预处理。
所述身份认证模块采用前述的指静脉识别方法中的检测模型,将预处理后的图像提取特征进行合法性判断,通过与存储模块备份的合法指静脉图像比较判断是否为合法用户。
所述存储模块存储有合法用户的指静脉图像和用户信息、指静脉识别方法以及新录入的待检测用户的指静脉图像。
所述结果输出模块包括合法通过单元和非法报警单元。合法通过单元如门禁等,非法报警单元如蜂鸣器、警示灯和或提示语音等。
存储介质
本发明还提供了一种存储介质,其上存储有计算机指令,所述计算机指令运行时执行前述指静脉识别方法的步骤。其中,所述方法请参见前述 部分的详细介绍,此处不再赘述。
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成,该程序可以存储于计算机可读存储介质中,计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括,但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算设备访问的信息。按照本文中的界定,计算机可读介质不包括暂存电脑可读媒体(transitory media),如调制的数据信号和载波。
本申请各部分操作所需的计算机程序编码可以用任意一种或多种程序语言编写,包括面向对象编程语言如Java、Scala、Smalltalk、Eiffel、JADE、Emerald、C++、C#、VB.NET、Python等,常规程序化编程语言如C语言、VisualBasic、Fortran2003、Perl、COBOL2002、PHP、ABAP,动态编程语言如Python、Ruby和Groovy,或其他编程语言等。该程序编码可以完全在用户计算机上运行、或作为独立的软件包在用户计算机上运行、或部分在用户计算机上运行部分在远程计算机运行、或完全在远程计算机或处理设备上运行。在后种情况下,远程计算机可以通过任何网络形式与用户计算机连接,比如局域网(LAN)或广域网(WAN),或连接至外部计 算机(例如通过因特网),或在云计算环境中,或作为服务使用如软件即服务(SaaS)。
最后应说明的是:以上实施例仅用以说明本发明的技术方案,而非对其限制;尽管参照前述实施例对本发明进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围。
Claims (9)
- 一种基于多方合作的指静脉识别方法,其特征在于,方法包括:S1、数据采集:采集指静脉图像,并对采集的图像进行数据标注和清洗,以此规范化训练数据集;S2、模型训练:首先建立引入注意力机制的训练模型,再基于多用户协同训练框架,多合作方之间在无数据交换的情况下共同训练得到优化的全局模型;S3、模型测试:以步骤S2优化后的全局模型作为检测模型,将测试数据集带入检测模型进行检测。
- 根据权利要求1所述的指静脉识别方法,其特征在于:在步骤S2中,初始建立的训练模型采用卷积神经网络结构,且每一部分之后都增加一个注意力机制,形成具有注意力机制的网络结构。
- 根据权利要求2所述的指静脉识别方法,其特征在于:所述卷积神经网络结构采用轻量网络的MobileNet或ShuffleNet算法模型,注意力机制采用包括空间注意力机制模块和通道注意力机制模块的CBAM注意力机制。
- 根据权利要求1所述的指静脉识别方法,其特征在于:步骤S2中优化的全局模型获得包括以下训练步骤:S221、合作方初始化模型参数;S222、合作方更新本地模型并使用本地数据进行训练,得到N个本地模型;S223、各个合作方分别对本地模型参数进行同态加密,将加密后的模型参数发送给合作中心;S224、合作中心将N个模型参数进行聚合得到全局模型,将全局模型参数返回给各个合作方;S225、各个合作方对返回的模型参数解密,重复S222~S224直至模型收敛或达到预设训练轮数,以此获得优化的全局模型。
- 根据权利要求4所述的指静脉识别方法,其特征在于:在步骤S223中,本地模型的参数通过同态加密算法进行加密。
- 根据权利要求5所述的指静脉识别方法,其特征在于:所述同态加密算法采用RAS算法,算法过程包括:a、产生公钥和私钥:随意选择两个大的质数p和q,p不等于q,令N=pq,r=(p-1)(q-1);随机选择一个小于r的整数e(满足gcd(e,r)=1),求得e关于r的模反元素d(即e×d=1mod r),则有公钥(N,e),私钥(N,d);b、明文m,加密为c:c=m e mod N;c、密文为c,解密为m:m=c d mod N。
- 根据权利要求1所述的指静脉识别方法,其特征在于:在步骤S3中测试数据集包括合法用户的手指静脉模型图集和非法用户的手指静脉模型图。
- 一种指静脉身份识别***,其特征在于:***包括指静脉图像采集模块、图像预处理模块、身份认证模块、存储模块和结果输出模块,所述指静脉图像采集模块采用指静脉采集仪,其与图像预处理模块和存储模块电信连接;所述图像预处理模块将预存在存储模块的指静脉图像进行预处理;所述身份认证模块采用权利要求1-7任一项所述的指静脉识别方法中的检测模型,将预处理后的图像提取特征进行合法性判断,通过与存储模块备份的合法指静脉图像比较判断是否为合法用户;所述存储模块存储有合法用户的指静脉图像和用户信息、指静脉识别方法以及新录入的待检测用户的指静脉图像;所述结果输出模块包括合法通过单元和非法报警单元。
- 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指令,所述计算机指令当运行时执行权利要求1-7任一项所述的指静脉识别方法。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111028126.6A CN113792632A (zh) | 2021-09-02 | 2021-09-02 | 基于多方合作的指静脉识别方法、***和存储介质 |
CN202111028126.6 | 2021-09-02 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023029274A1 true WO2023029274A1 (zh) | 2023-03-09 |
Family
ID=79182571
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/136425 WO2023029274A1 (zh) | 2021-09-02 | 2021-12-08 | 基于多方合作的指静脉识别方法、***和存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN113792632A (zh) |
WO (1) | WO2023029274A1 (zh) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
CN117274824A (zh) * | 2023-11-21 | 2023-12-22 | 岭南设计集团有限公司 | 一种基于人工智能的红树林生长状态检测方法及*** |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107967462A (zh) * | 2017-12-12 | 2018-04-27 | 天津津航计算技术研究所 | 一种基于指静脉生物特征的身份识别方法 |
US20190340344A1 (en) * | 2016-11-18 | 2019-11-07 | All It Top Co., Ltd. | Input/output integration module for simultaneously linking biological information algorithms |
CN112949570A (zh) * | 2021-03-26 | 2021-06-11 | 长春工业大学 | 一种基于残差注意力机制的指静脉识别方法 |
CN113255759A (zh) * | 2021-05-20 | 2021-08-13 | 广州广电运通金融电子股份有限公司 | 基于注意力机制的目标内特征检测***、方法和存储介质 |
CN113255210A (zh) * | 2021-05-13 | 2021-08-13 | 燕山大学 | 一种风电机组联邦故障诊断方法及*** |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110147732A (zh) * | 2019-04-16 | 2019-08-20 | 平安科技(深圳)有限公司 | 指静脉识别方法、装置、计算机设备及存储介质 |
CN110442457A (zh) * | 2019-08-12 | 2019-11-12 | 北京大学深圳研究生院 | 基于联邦学习的模型训练方法、装置及服务器 |
CN112001292B (zh) * | 2020-08-18 | 2024-01-09 | 大连海事大学 | 一种基于多尺度注意力机制深度哈希的指静脉索引方法 |
CN111931876B (zh) * | 2020-10-12 | 2021-02-05 | 支付宝(杭州)信息技术有限公司 | 一种用于分布式模型训练的目标数据方筛选方法及*** |
CN112668472B (zh) * | 2020-12-28 | 2021-08-31 | 中国科学院自动化研究所 | 基于联邦学习的虹膜图像特征提取方法、***和装置 |
CN113312989B (zh) * | 2021-05-11 | 2023-06-20 | 华南理工大学 | 一种基于聚合描述子与注意力的指静脉特征提取网络 |
-
2021
- 2021-09-02 CN CN202111028126.6A patent/CN113792632A/zh active Pending
- 2021-12-08 WO PCT/CN2021/136425 patent/WO2023029274A1/zh unknown
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190340344A1 (en) * | 2016-11-18 | 2019-11-07 | All It Top Co., Ltd. | Input/output integration module for simultaneously linking biological information algorithms |
CN107967462A (zh) * | 2017-12-12 | 2018-04-27 | 天津津航计算技术研究所 | 一种基于指静脉生物特征的身份识别方法 |
CN112949570A (zh) * | 2021-03-26 | 2021-06-11 | 长春工业大学 | 一种基于残差注意力机制的指静脉识别方法 |
CN113255210A (zh) * | 2021-05-13 | 2021-08-13 | 燕山大学 | 一种风电机组联邦故障诊断方法及*** |
CN113255759A (zh) * | 2021-05-20 | 2021-08-13 | 广州广电运通金融电子股份有限公司 | 基于注意力机制的目标内特征检测***、方法和存储介质 |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
CN117274824A (zh) * | 2023-11-21 | 2023-12-22 | 岭南设计集团有限公司 | 一种基于人工智能的红树林生长状态检测方法及*** |
CN117274824B (zh) * | 2023-11-21 | 2024-02-27 | 岭南设计集团有限公司 | 一种基于人工智能的红树林生长状态检测方法及*** |
Also Published As
Publication number | Publication date |
---|---|
CN113792632A (zh) | 2021-12-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023029274A1 (zh) | 基于多方合作的指静脉识别方法、***和存储介质 | |
Tariq et al. | A review of deep learning security and privacy defensive techniques | |
Kumar et al. | Privacy preserving security using biometrics in cloud computing | |
US20190207960A1 (en) | Detecting network attacks | |
CN1860724B (zh) | 用于标识的方法 | |
US11354797B2 (en) | Method, device, and system for testing an image | |
EP4105821A1 (en) | Face liveness detection method, system and apparatus, computer device, and storage medium | |
CN105681324A (zh) | 互联网金融交易***及方法 | |
CN117390657A (zh) | 数据加密方法、装置、计算机设备和存储介质 | |
US20210173910A1 (en) | Thermally enriched multi-modal and multi-channel biometric authentication | |
US20210044659A1 (en) | Adaptive cross-channel tracking of electronic records signature modifications | |
Wu et al. | Fingerprint bio‐key generation based on a deep neural network | |
CN113055153B (zh) | 一种基于全同态加密算法的数据加密方法、***和介质 | |
Dahanayaka et al. | Robust open-set classification for encrypted traffic fingerprinting | |
Li et al. | A four-dimensional space-based data multi-embedding mechanism for network services | |
US11436340B2 (en) | Encrypted device identification stream generator for secure interaction authentication | |
Subash et al. | Real-time behavioral biometric information security system for assessment fraud detection | |
Xiong et al. | Detecting GAN-based privacy attack in distributed learning | |
US20230344828A1 (en) | System and method for intelligent authentication via object movement recognition | |
Duan et al. | SSGD: A safe and efficient method of gradient descent | |
Chang et al. | [Retracted] Application of Face Recognition in E‐commerce Security Authentication in the Era of Big Data | |
CN111275603B (zh) | 一种基于风格转换的安全图像隐写方法与电子装置 | |
CN115964706A (zh) | 联邦学习场景下的训练数据毒化防御方法 | |
Yao et al. | A novel privacy-preserving neural network computing approach for E-Health information system | |
Zhang et al. | Construct new graphs using information bottleneck against property inference attacks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
NENP | Non-entry into the national phase |
Ref country code: DE |