Biometric authentication method based on adaptive loss function, medium and device

Abstract

The present invention provides a biometric authentication method, medium and device based on an adaptive loss function; wherein the method comprises: using a neural network model to extract features from information to be authenticated, and then obtaining a biometric authentication result based on the extracted features; when training the neural network model, an adaptive loss function is used: wherein, θ l The adaptive loss function used in this method has very stable and superior performance, thereby improving the performance of the neural network model and the accuracy of biometric authentication.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biometric authentication, more particularly, to a biometric authentication method based on an adaptive loss function, a medium and an apparatus. BACKGROUND

[0002] Biometric authentication refers to the science and technology of realizing automatic identity authentication by acquiring and analyzing physiological characteristics (such as face, fingerprint, iris, palm print, and vein) and / or behavioral characteristics (voiceprint, gait, and signature) of human body. With the rapid development of the Internet and the high-speed promotion of intelligentization, biometric authentication technology has penetrated into all aspects of people's production and life, from electronic device unlocking, supermarket cash register, community access control, to high-speed rail station and airport security check. Biometric characteristics have become an important digital identity credential for people in the era of Internet of Everything, and biometric authentication has become the preferred solution to solve identity security problems ("end password") in the process of current social intelligentization development. Although the biometric authentication method has higher security than the traditional authentication methods such as password and magnetic card, however, with the development of technology, the problems faced by biometric authentication technology in security are increasingly prominent. In order to further improve the security of the authentication system, the performance of the identity feature extractor is crucial.

[0003] When designing a feature extractor, the loss function directly determines the optimization direction of the feature extractor, and thus plays a decisive role in the performance of the final feature extractor. Current loss functions can be roughly divided into two categories: metric learning-based loss functions and classification-based loss functions. Among the metric learning-based loss functions, contrastive loss and triplet loss are two typical loss functions. Contrastive loss first pairs the samples and reduces the distance between positive sample pairs and increases the distance between negative sample pairs through training. Triplet loss further combines three samples, including two samples of the same class and one sample of a different class, and then trains the model to separate the two matching samples from the sample of a different class and form a certain interval. Subsequent metric learning-based loss functions are basically designed based on contrastive loss and triplet loss. However, since this type of method involves sample combination and difficult sample mining problems, the training is relatively inefficient and the performance of the obtained model needs to be improved. Classification-based loss functions are basically based on the Softmax function (cross-entropy loss). These loss functions are proxy-based methods. The proxy refers to the weight matrix in Softmax, where each row of the matrix represents a cluster center of a class. In classification-based loss functions, NormFace analyzes the importance of feature and weight normalization and replaces the inner product similarity function with a cosine similarity function, resulting in a significant performance improvement, so subsequent classification-based loss functions are basically based on the improvement of NormFace. On the basis of NormFace, SphereFace, CosFace, ArcFace, and AmpFace respectively introduce multiplicative angular margin, additive cosine margin, additive angular margin, and multiplicative cosine margin, thereby further improving the authentication accuracy of the model. Considering that the above margin coefficients are fixed, it is difficult to meet the needs of different samples, as different samples come from different users or are collected in different environments. To solve the above problem, some adaptive margin methods are proposed. Among them, AdaptiveFace proposes a method of assigning a learnable margin parameter to each class, Dyn-arcFace designs a margin parameter allocation method based on the inter-class margin, and MV-Softmax and MagFace respectively allocate margin parameters according to the difficulty and quality of the samples. However, the above methods are not sufficient and intuitive in allocating margin parameters. SUMMARY

[0004] In order to overcome the defects and deficiencies in the prior art, the purpose of the present application is to provide a biological feature authentication method, medium and equipment based on an adaptive loss function; the adaptive loss function used in the method has very stable and superior performance, thereby improving the performance of the neural network model and improving the accuracy of biological feature authentication.

[0005] In order to achieve the above-mentioned purpose, the present application is implemented by the following technical solutions: a biological feature authentication method based on an adaptive loss function, which adopts a neural network model to extract features from to-be-authenticated information, and then obtains a biological authentication result according to the extracted features;

[0006] When the neural network model is trained, an adaptive loss function is used Adaptive loss function is:

[0007]

[0008] Wherein, θ l represents the angle between the agent and the normalized sample feature ; x represents the sample feature input to the Softmax function of the neural network model, l represents the label of the sample feature x; W l represents the lth row of the Softmax function weight matrix of the neural network model, s represents a scale coefficient; C is the number of training sample categories; a is a constant; and m represents an adaptive interval coefficient;

[0009] The adaptive interval coefficient m is calculated by the following formula:

[0010] m = max (m a + Δm, m th )

[0011]

[0012] Wherein, m a represents the current interval; m th represents the minimum value of the set adaptive interval coefficient m; Δm represents the interval difference value; clamp represents the truncation operation, if out of range, it is truncated to D min or D max ; m u represents the upper limit of interval exploration.

[0013] Preferably, the calculation method of the current interval m a is: the normalized sample feature As a boundary, the two-dimensional space about the agent corresponds to the existence of a mirror boundary The rows of the weight matrix of the Softmax function of the neural network model are regarded as agents, and the boundaries and mirror boundaries are counted The angle set alpha of all agents except the agent , and the angle set beta between all agents

[0014] Calculate the current interval m a :

[0015] m a = min(alpha)-theta l = min(beta)-2theta l .

[0016] Preferably, in the calculation formula of the adaptive interval coefficient m, D min = 0; D max = 0.1.

[0017] Preferably, the calculation process of the adaptive interval coefficient m performs gradient backpropagation blocking operation.

[0018] A readable storage medium, wherein the storage medium stores a computer program, the computer program causes the processor to execute the biological feature authentication method based on the adaptive loss function when the processor executes.

[0019] A computer device, comprising a processor and a memory for storing a processor executable program, when the processor executes the program stored in the memory, the biological feature authentication method based on the adaptive loss function is realized.

[0020] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0021] 1、The adaptive loss function in the present application adopts a linear similarity measure function, which can make the gradient value of the loss function monotonically decrease with the model quality, so as to quickly converge in the early training and finely adjust the model in the later training;

[0022] 2、The adaptive interval method proposed in the present application can assign a customized interval coefficient to the sample according to the sample separability and the model quality at each iteration; using this method can constrain the gradient value size, avoid excessively large gradient values and provide appropriate size gradient values for training, so that the training process is more stable;

[0023] ​3、The linear similarity measure function and the adaptive interval method are complementary in the present application, and the adaptive loss function obtained by combining the two methods has very stable and superior performance, thereby improving the performance of the neural network model and improving the accuracy of biometric authentication. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1(a) and Fig. 1(b) are modulation gradient value curves, respectively;

[0025] Fig. 2(a) and Fig. 2(b) are current interval definitions and distributions, respectively. DETAILED DESCRIPTION

[0026] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0027] Embodiment One

[0028] The biometric authentication method based on an adaptive loss function in the present embodiment adopts a neural network model to perform feature extraction on to-be-authenticated information, and then obtains a biometric authentication result according to the extracted features; during training of the neural network model, an adaptive loss function is adopted Adaptive loss function is as follows:

[0029]

[0030] wherein, represents the angle between the agent l and the normalized sample feature ; x represents the sample feature input to the Softmax function of the neural network model, l represents the label of the sample feature x; W l represents the lth row of the weight matrix of the Softmax function of the neural network model, s represents a scale coefficient; C is the number of training sample categories; a is a constant; and m represents an adaptive interval coefficient. The adaptive interval coefficient m is calculated by the following formula:

[0031] m = max(m a + Δm, m th )

[0032] m = max(m a + Δm, m th )

[0033]

[0034] wherein, m a represents the current interval; and m threpresents the minimum value of the set adaptive margin coefficient m; represents the margin difference value; clamp represents the truncation operation, if out of range, then truncate to D min or D max ; m u represents the upper limit of the margin exploration. Preferably, D min = 0; D max = 0.1.

[0035] The calculation process of the adaptive margin coefficient m is gradient blocking operation.

[0036] The adaptive loss function disclosed by the application is an improvement on the Softmax loss function based on the margin coefficient. Specifically, the adaptive loss function of the application first proposes to replace the cosine function commonly used in the current loss function with a linear function. By using a linear function, the main gradient value of the loss function will monotonically decrease as the model quality improves, so that the model can be quickly optimized in the early training and fine-tuned in the late period. On the basis of the linear function, the application further proposes an adaptive margin method to adaptively assign a dedicated margin coefficient to each sample according to the separability of the sample and the quality of the current model. The adaptive method can constrain the gradient value, thereby reducing the excessive gradient and generating an appropriate gradient value, thereby achieving more stable model training. The proposed linear function and adaptive margin method are complementary. Combined with both, a more high-quality biometric authentication model can be trained.

[0037] The design idea of the adaptive loss function of the application is:

[0038] First, the existing Softmax loss function based on the margin coefficient is defined as follows:

[0039]

[0040] wherein, W j represents the lth row of the Softmax weight matrix x represents the sample feature input to the Softmax l represents the label of the sample x, θ l represents the angle between and , s represents the scale coefficient, and C is the number of training sample categories. Each row of W represents a proxy corresponding to a cluster center of a category. On this basis, after introducing the margin coefficient, formula 1 becomes:

[0041]

[0042] where m0, m1, m2 and m3 represent the multiplicative cosine spacing coefficient (AmpFace), the multiplicative angle spacing coefficient (SphereFace), the additive angle spacing coefficient (ArcFace), and the additive cosine spacing coefficient (CosFace), respectively. In actual use, one or more spacing coefficients can be used.

[0043] Next, the similarity measure function is designed:

[0044] The current mainstream Softmax loss function based on spacing coefficients is designed by geometric analysis. Considering that these loss functions all have reasonable and similar geometric interpretations, but the performance is different, the present application analyzes the differences in gradients and seeks improvement. The gradient of the loss function with respect to W is:

[0045]

[0046]

[0047] The gradient value of the corresponding agent W l and the non-corresponding agent W o can be simplified as:

[0048]

[0049]

[0050] where y l is given in equation (2), and y o is as follows:

[0051]

[0052] In equations (5) and (6), the gradient size includes two main parts: the basic gradient value B and the modulation gradient value M, which are defined as follows:

[0053] B l = |y l -1|, B o = |y o -0|, (8)

[0054] M l = |m0m1sin(m1θ l +m2)|, M o = |sin(θ o )| (9)

[0055] In training, the present application expects θ l to be constantly smaller, and θo can keep increasing, while y l can keep approaching 1, y o can keep approaching 0, and finally B l and B o can keep approaching 0. It can be seen that B l is a monotonically increasing function of θ l , and B o is a monotonically decreasing function of θ o . This monotonicity is desirable because it can make the model obtain a larger gradient value at the beginning of training, thereby accelerating the training speed, and obtain a smaller gradient at the end of training, thereby fine-tuning the model.

[0056] For the modulated gradient M, the present application plots four typical cases in Fig. 1(a) and Fig. 1(b), which are SphereFace (m0=1, m1=1.2, m2=0), CosFace (m0=1, m1=1, m2=0), ArcFace (m0=1, m1=1, m2=0.5) and AmpFace (m0=0.35, m1=1, m2=0) respectively, where m3 has no effect on the result (refer to formula (9)). Since B l is a monotonically increasing function of θ l , and B o is a monotonically decreasing function of θ o , M l should be a non-monotonically decreasing function of θ l , and M o should be a non-monotonically increasing function of θ o , however, according to Fig. 1(a) and Fig. 1(b), the properties of M l and M o do not meet the requirements. Therefore, the monotonicity of the main gradient value (B l M l and B o M o ) of the Softmax function based on the margin coefficient will become undesirable under the modulation of M. In order to improve the monotonicity of the main gradient value of the Softmax function based on the margin coefficient, the present application proposes to remove the modulation of M on B. Through analysis, it is equivalent to:

[0057]

[0058] Through calculation, we can get:

[0059]

[0060] where p is an arbitrary constant. Therefore, the Softmax function based on the margin coefficient is improved as:

[0061]

[0062] where a is a constant, and m is an introduced interval coefficient (a+m is also a constant).

[0063] The proposed linear similarity measure function has the ideal monotonicity of gradient value in addition to the gradient value. Since the similarity measure function is linear, there is no need to handle the boundary problem. The loss function based on the cosine function needs to handle the problem of angle plus interval coefficient exceeding the specified monotonic interval ([0, π]).

[0064] Then, an adaptive interval method is used:

[0065] The linear function only ensures that the monotonicity of the gradient value meets the training requirements (the gradient value monotonically decreases as the model quality improves). The size of the gradient value cannot be too large or too small. If the size of the gradient value is not constrained, some low-quality samples, difficult samples, and samples with labeling errors will have a greater gradient impact on model optimization. When the loss value is close to the global minimum value, the gradient value that is too large will cause the loss value to oscillate or diverge near the minimum value. Therefore, the present application proposes an adaptive interval method to enhance the proposed linear function-based Softmax loss function (equation (12)). Since the gradient value of the linear function-based Softmax loss function is mainly related to the basic gradient value B (the modulation gradient value is 1), the present application further analyzes y l and y o and makes the following transformation:

[0066]

[0067] Since the difference between θ l and θ c (or θ o ) is large, the difference between θ c and θ o is small, so (θ l -θ c +m) and (θ o -θ l -m) have an important influence on the size of y l and y o . It can be seen that the value of m is very important. If m is small, a larger interval cannot be obtained, and on the contrary, it will lead to the generation of excessive gradient.

[0068] To reasonably allocate the margin coefficient m, the present application first proposes the concept of current margin. The current margin can be used to measure the separability of samples and the quality of the model. In order to facilitate understanding without losing generality, the present application takes a three-classification problem in two-dimensional space as an example for explanation, and the relevant concepts and conclusions are also applicable to multi-classification problems in high-dimensional space. As shown in FIG. 2(a), are the proxies of classes l, i, j respectively, i.e. three row vectors of the Softmax weight matrix. is the feature vector of the sample of class l, so should be as close as possible to its corresponding proxy (i.e. reduce the angle θ l ) and far away from non-corresponding proxies and (i.e. increase the angles θ i and θ j ).

[0069] The most reasonable and intuitive way to allocate the margin coefficient m is to set it according to the real margin. However, the real margin is a global statistical quantity that needs to be calculated according to the distribution of all samples. Specifically, the real margin is equal to the inter-class distance (i.e. β i and β j ) minus twice the minimum value of the angle from the boundary to its corresponding proxy. In high-dimensional space, the boundary of each class is the hyper-conic surface formed by the sample feature with the largest angle distance from the proxy of all sample features of the class. Since it is difficult to calculate the real margin at each iteration during actual training, the present application proposes the current margin to approximate the real margin.

[0070] As shown in FIG. 2(a), when calculating the current margin, each normalized sample feature is regarded as a boundary, and correspondingly, there is a mirror boundary symmetric to the corresponding proxy. In three-dimensional space, the boundary is which rotates around its corresponding proxy to form a conic surface. According to the definition of the boundary and the margin, θ l +m j = α j , so m j = α j - θ l = β j - 2θ l . Similarly, m i can be obtained according to the mirror boundary of , where α i and α j are the angles from the boundary to non-corresponding proxies and . Finally, mi and m j The minimum value in is defined as the current interval (the minimum value of the current interval of all samples is the true interval):

[0071] m a =min(α)-θ k =min(β)-2θ l (14)

[0072] Among them, α represents the boundary and mirror boundaries To remove the agent The angle set of all agents except ; β represents the set of angles between all agents.

[0073] The current interval can well represent the current separability of the sample and the quality of the model. Figure 2(b) shows the distribution of the current interval obtained from the SCUT-DHGA gesture authentication dataset. It can be seen that the current interval will change significantly with different users and different samples. Therefore, using a fixed interval may allocate too large a gradient when the sample has a small current interval and produce an insufficient gradient when the sample has a large current interval. Assume that the minimum angle between the boundary and the non-corresponding agent is α m (α m =min(α)), then formula (13) can be further converted to:

[0074]

[0075] Where Δm=mm a (m a =α m -θ l ). Therefore, according to the current interval m a The allocation interval coefficient m can more appropriately adjust y l and y o Then obtain a moderate gradient value.

[0076] Considering that samples have different qualities and difficulties, and may even be incorrectly labeled, the present invention proposes to progressively assign an interval coefficient m to samples based on the current interval, especially when the samples have a certain degree of separability. The adaptive interval coefficient is defined as follows:

[0077]

[0078] where m u Indicates the upper limit of interval exploration (when m a More than m u When the method of the present invention stops allocating more than m a The interval coefficient m is used for model training), clamp represents the truncation operation, is limited between 0 and 0.1. If it is out of the range, it is truncated to 0 or 0.1. Δm only works when m a + Δm > m th . In the case of m a + Δm ≤ m th , m = m th . In this case, the model quality is low, so the present invention sets the minimum value of the interval coefficient m th to avoid the optimization speed too slow in the early training. When Δm works, the model already has a higher quality and the samples also have higher separability, because the current interval is already large (m a > m th - Δm and max(Δm) = 0.1). In order to further improve the model quality and sample separability, the present invention sets the maximum value of Δm to 0.1 to continue to expand the current interval, limiting the maximum value to 0.1 can effectively reduce the loss value shock or divergence, so that the training is more stable. The present invention sets the minimum value of Δm to 0, so that when m a > m u , the size of the current interval is kept as large as possible. In combination with the proposed linear similarity measure function and the adaptive interval method, the final adaptive loss function is defined as follows:

[0079]

[0080] It should be noted that the present invention performs gradient backpropagation blocking operation on the calculation process of the adaptive interval m during training. Because if not blocked, there will be a pathological problem, that is, to reduce the loss value by reducing the adaptive interval m, instead of reducing θ l and increasing θ o .

[0081] The advantages and positive effects of the method of the present invention are:

[0082] 1.The linear similarity measure function can make the gradient value of the loss function monotonically decrease with the model quality, so as to quickly converge in the early training and finely tune the model in the later training. When the linear similarity measure function and the fixed interval coefficient are applied to the Softmax loss function, the present application obtains a loss function based on an additive angular interval coefficient, i.e. formula (12). Considering that ArcFace and QamFace are both loss functions based on an additive angular interval, and ArcFace uses a cosine function for similarity measure and QamFace uses a quadratic function for similarity measure. Therefore, comparing the adaptive loss function of the present application with ArcFace and QamFace can better prove the effect of the linear similarity measure function. (Note that the adaptive interval method proposed in the present application is not used at this time, but the traditional fixed interval coefficient is used). Since the loss function involves two hyperparameters s and m. Grid search is performed on the SCUT-DHGA gesture authentication dataset, and 12 groups of search are performed for each loss function. The adaptive loss function of the present application is denoted as L3AM (Linear Adaptive Additive Angular Margin Loss), and the equal error rate (the lower the better) of the top 5 experiments with the best performance is shown in Table 1.

[0083] Table 1 Equal error rate of the top 5 experiments with the best performance

[0084] Loss function L3 AM (%) ArcFace (%) QamFace (%) 1 0.611 0.652 0.630 2 0.622 0.696 0.670 3 0.630 0.719 0.670 4 0.648 0.744 0.693 5 0.682 0.763 0.707

[0085] 2.The proposed adaptive interval method can assign a customized interval coefficient to the sample at each iteration according to the sample separability and model quality. Using this method can constrain the gradient value, avoid excessively large gradient values and provide appropriate size gradient values for training, so as to make the training process more stable. We conducted a comparative experiment on the SCUT-DHGA gesture authentication dataset, and the results are shown in Table 2. It is found that the adaptive interval method can further reduce the equal error rate.

[0086] Table 2 Comparative experiment results on the SCUT-DHGA gesture authentication dataset

[0087]

[0088] 3.The proposed linear similarity measure function and adaptive interval method are complementary in advantages, and the final L3AM loss function obtained by combining the two methods has very stable and superior performance. On the SCUT-DHGA gesture authentication dataset, it can achieve better performance than the current 25 mainstream loss functions. In addition, it also performs well on 6 palmprint and palm vein datasets.

[0089] 25 loss function performance results are: ArcFace: 0.652%; Dyn-arcFace: 0.641%; ElasticFace-Arc: 0.722%; CurricularFace: 0.626%; AdaptiveFace: 0.648%; RegularFace: 0.882%; MV-Softmax: 0.730%; MagFace: 1.004%; IHEM: 0.634%; Regss: 0.737%; Softmax: 1.359; CenterLoss: 1.771%; NormFace: 0.804%; AdaCos: 1.526%; EqMLoss: 0.730%; CVMLoss: 0.659%; SFace: 0.734%; CircleLoss: 0.844%; CosFace: 0.689%; SphereFace: 0.704%; SphereFace2: 1.030%; SphereFace+: 0.678%; SphereFaceR: 0.641%; DSoftmax: 0.974%; L3AM: 0.559%.

[0090] The comparison results on six palmprint and palm vein datasets (CASIA, PolyU and Tongji three datasets contain both palmprint and palm vein data, so there are 6 datasets in total) are shown in Table 3.

[0091] Table 3 Comparison results on six palmprint and palm vein datasets

[0092]

[0093] Embodiment Two

[0094] The embodiment provides a readable storage medium, wherein the readable storage medium stores a computer program, and the computer program causes a processor to execute the method in the embodiment one when the computer program is executed by the processor.

[0095] Embodiment Three

[0096] The embodiment provides a computer device, which comprises a processor and a memory for storing a program executable by the processor, and the processor implements the method in the embodiment one when executing the program stored in the memory.

[0097] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations and simplifications made without departing from the spirit and principle of the present application should be equivalent replacement methods, and all are included in the protection scope of the present application.

Claims

1. A biometric authentication method based on an adaptive loss function, characterized by: A neural network model is used to extract features from the information to be authenticated, and then a biometric authentication result is obtained based on the extracted features; the information to be authenticated is one of the following: face, fingerprint, iris, gesture, palm print, and vein; When the neural network model is trained, an adaptive loss function is used. Adaptive loss function for: Among them, θ l Representation Agent and normalized sample features The angle between x represents the sample features input to the Softmax function of the neural network model, l represents the label of sample feature x; Each row of W represents an agent, corresponding to the cluster center of a category, W l Represents the lth row of the weight matrix of the Softmax function of the neural network model, s represents the scale coefficient; C is the number of training sample categories; a is a constant; m represents the adaptive interval coefficient; The adaptive interval coefficient m is calculated by the following formula: m=max(m a +Δm,m th ) Among them, m a Indicates the current interval; m th Indicates the minimum value of the adaptive interval coefficient m; Δm indicates the interval difference; clamp indicates the truncation operation. If out of range, truncate to D min or D max ;m u represents the upper limit of interval exploration; The current interval m a The calculation method is: each normalized sample feature As a boundary, the two-dimensional space is about the agent There is a mirror boundary corresponding to the symmetric position of The row vectors of the weight matrix of the Softmax function of the neural network model are used as agents to statistically analyze the boundary and mirror boundaries To remove the agent The set of angles α of all agents other than , and the set of angles between all agents β; Calculate the current interval m a : m a =min(α)-θ l =min(β)-2θ l 。 2. The biometric authentication method based on an adaptive loss function according to claim 1, wherein: In the calculation formula of the adaptive interval coefficient m, D min =0;D max =0.

1.

3. The biometric authentication method based on adaptive loss function according to claim 1, characterized in that: The calculation process of the adaptive interval coefficient m performs a gradient backpropagation blocking operation.

4. A readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, enables the processor to perform the biometric authentication method based on an adaptive loss function according to any one of claims 1 to 3.

5. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the biometric authentication method based on the adaptive loss function according to any one of claims 1 to 3 is implemented.

Images