Joint method and system for face recognition and image quality assessment based on adaptive mechanism

Abstract

The present invention provides a combined method and system for face recognition and image quality assessment using an adaptive mechanism, including: constructing an adaptive scaling factor related to feature amplitude; constructing an angle penalty to form a joint adaptive mechanism with the scaling factor; constructing a regularization term that is adaptive to feature amplitude; utilizing the scaling factor, angle penalty, and regularization term to form a cross-entropy loss function in an angular space for face recognition and face image quality assessment; constructing a face feature extractor to extract facial features and obtain feature amplitudes; constructing a face identity determinator to predict identities based on facial features; and training the face feature extractor and face identity determinator based on the cross-entropy loss function. The present invention proposes a joint adaptive mechanism of the scaling factor and angle penalty to optimize the distribution of feature amplitudes so that they approximately follow a normal distribution, thereby learning more discriminative facial features and obtaining a face image quality score that is more consistent with the recognition results.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a combined method, system, medium and terminal for face recognition and image quality assessment with an adaptive mechanism. Background Art

[0002] Face recognition is a key task in intelligent image processing and has been a research hotspot in the field of computer vision in recent years. It is applicable to technical fields such as identity verification and public security. Face recognition can be defined as extracting facial features from a facial image and then using these features to confirm or verify the identity corresponding to the image in an identity database. It is an automated identity verification technology that can quickly confirm the identity of a target person based on an input facial image within an identity database. It is a critical step in applications such as identity security verification and intelligent video surveillance.

[0003] A major challenge facing the practical application of face recognition systems in open environments is the quality of facial images. Compared to the standardized, aligned facial images in datasets, the quality of facial images in real-world applications varies greatly due to acquisition conditions (such as lighting, shooting angle, and lens focal length). Facial image quality is closely related to the accuracy of identity recognition. Therefore, it is imperative to incorporate an image quality assessment module into face recognition systems. This module can effectively eliminate facial images that do not meet recognition standards, thereby improving recognition reliability. Consequently, research on facial image quality assessment has garnered increasing attention in recent years.

[0004] From the existing technical literature search, it is found that there are two main research ideas for face image quality assessment methods. One idea is the image-based method, which emphasizes the consistency with the subjective quality evaluation of the human eye and extracts image parameters from different angles to evaluate the quality, including portrait characteristics (such as lighting, posture, expression, etc.) and image attributes (camera focal length, image resolution, etc.) (see Matteo Ferrara, Annalisa Franco, Dario Maio, and Davide Maltoni. Face image conformance to iso / icao standards in machine readable travel documents. IEEE Transactions on Information Forensics and Security, 7(4):1204-1213, 2012; Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics (ICB), pages 1-8, 2019.); However, image-based methods can only focus on a limited number of factors consistent with human visual assessment, and these factors that affect image quality are not always effective for robotic face recognition systems. Another approach is identity-based methods, which focus more on the consistency of quality assessment with machine recognition. Most of these methods use deep learning-based methods, using convolutional neural networks (CNNs) to extract facial image features and assess the identity-based quality of facial images, improving the generalization ability of various face recognition methods.(See: Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embeddingrobustness. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020; Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, YugeHuang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7670-7679, June 2021.). However, all of the above methods work independently of the face recognition method. In actual applications in open environments, they will increase the computing power overhead of the face recognition system.

[0005] Recent research is focused on achieving an end-to-end joint approach for face recognition and quality assessment, offering a solution that optimizes the distribution of facial features within the feature space. (See: Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In Proceedings of the IEEE / CVF International Conference on Computer Vision (ICCV), October 2019; Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14225-14234, June 2021.) However, the quality score distribution generated by this approach is overly concentrated and not normal. According to the central limit theorem, the quality of large-scale face images in the wild should be close to a normal distribution. Therefore, the quality distribution characteristics of existing joint face recognition and quality assessment methods affect the reliability and generalization of the quality scores, and also affect the performance of recognition tasks. Summary of the Invention

[0006] In view of the defects in the prior art, the purpose of the present invention is to provide a combined method and system for face recognition and image quality assessment with an adaptive mechanism.

[0007] According to one aspect of the present invention, a method for combining face recognition and face image quality assessment using an adaptive mechanism is provided, comprising:

[0008] Constructing an adaptive scale factor related to the feature amplitude;

[0009] Constructing an angle penalty and forming a joint adaptive mechanism with the scale factor;

[0010] Construct a regularization term that is adaptive to the feature amplitude;

[0011] Using the scale factor, angle penalty and regularization term, a cross entropy loss function in angle space is formed for face recognition and face image quality assessment;

[0012] Constructing a face feature extractor to extract face features and obtain feature amplitudes;

[0013] constructing a face identity determiner that obtains a predicted identity based on the face features;

[0014] The facial feature extractor and facial identity determiner are trained based on the cross entropy loss function, and the trained facial feature extractor and facial identity determiner are used to complete face recognition, face identity determination and face image quality evaluation.

[0015] Preferably, the adaptive scale factor is s(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i The linear function of

[0016]

[0017] Among them, u a and l a are characteristic amplitudes a i The upper and lower bounds of u s and l s s(a i )’s upper and lower bounds.

[0018] Preferably, the angle penalty is m(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i A monotonically increasing convex function of , where:

[0019] m(a i )'s upper and lower bounds u m and l m satisfy

[0020] m(a i ) satisfies the first-order derivative of m'(a i )∈[k,K], k satisfies K is the upper bound of the angle penalty derivative function m'(a).

[0021] Preferably, the regularization term is g(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i Strictly monotone decreasing convex function, g(a i ) satisfies the first-order derivative of g'(u a )=0,u a and l a A i The upper and lower bounds of .

[0022] Preferably, the scale factor, angle penalty and regularization term are used to form a cross entropy loss function in the angle space for face recognition and face image quality evaluation:

[0023] The cross entropy loss function Any one of That is, for a single face image p i The loss function is:

[0024]

[0025] Among them, λ satisfies s(a i ) is the adaptive scale factor; m(a i ) is the angle penalty; g(a i ) is a regularization term; s is a fixed scale factor, and its value is u s ;

[0026] N represents the batch size, i represents the i-th face sample in the batch, x i Represents facial features, y i Represents x i Belongs to the identity category, θ j represents the j-th identity category weight W j with x i The angle between them, that is, the category distance, where j = 1, 2, 3, ..., C (j ≠ y i ), C is the total number of identities.

[0027] Preferably, a neural network FNN is designed as a face feature extractor FNN;

[0028] Design a neural network INN as a face identity determiner INN;

[0029] The face feature extractor FNN and the face identity determiner INN use the loss function to convert the identity corresponding to the face image, the amplitude of the face feature and the category distance As input, iterative optimization is performed.

[0030] Preferably, the loss function It is defined in [l a ,u a ] on a i A strictly convex function with a unique minimum point a * ;a * About intra-class distance Monotonically decreasing, a * About the distance between classes Monotonically increasing.

[0031] According to a second aspect of the present invention, a combined system for face recognition and face image quality assessment with an adaptive mechanism is provided, comprising:

[0032] A facial feature extraction module uses a facial feature extractor trained using the combined face recognition and face image quality assessment method. For any face image, the facial feature extractor extracts its facial features, and the magnitude of the facial features is the quality score of the face image.

[0033] A face identity determination module, which uses a face identity determiner trained using the combined face recognition and face image quality assessment method, to predict the corresponding identity for any facial feature.

[0034] The facial feature extraction module and the facial identity determination module complete facial recognition, facial identity determination and facial image quality evaluation.

[0035] According to a third aspect of the present invention, a joint terminal for face recognition and face image quality assessment is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor can be used to execute any one of the methods described, or to run the system, when executing the program.

[0036] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used for any of the methods described above, or for running the system described above.

[0037] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0038] 1. The combined method and system of face recognition and face image quality assessment with an adaptive mechanism in the embodiment of the present invention adopts a combined adaptive mechanism of scale factor and angle penalty to optimize the distribution of feature amplitudes so that it approximately obeys the normal distribution, thereby learning more discriminative facial features.

[0039] 2. The combined method and system of face recognition and face image quality assessment with an adaptive mechanism in the embodiment of the present invention adopts an adaptive mechanism (property of the loss function) in which the feature amplitude monotonically decreases with respect to the intra-class distance of the sample and monotonically increases with respect to the inter-class distance, to ensure that the feature amplitude after iterative optimization can represent the quality of the face image, thereby obtaining a face quality score that is more consistent with recognition.

[0040] 3. The combined method and system of face recognition and face image quality assessment with an adaptive mechanism in the embodiment of the present invention adopts a loss function that adapts to the self-feature amplitude and uses it to iteratively optimize the face feature extractor and identity determiner, so that the distribution of face features in the feature space is more discriminative and consistent, thereby obtaining more recognizable face features, more accurate identity recognition, and quality assessment that is more consistent with recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0042] Figure 1 A schematic diagram of the structure of a joint system for face recognition and face image quality assessment using an adaptive mechanism in a preferred embodiment of the present invention;

[0043] Figure 2 This is a comparison curve of face image quality evaluation in a preferred embodiment of the present invention and other advanced methods in the academic community;

[0044] Figure 3 Schematic diagram of the comparison results of the distribution of face image quality scores in a preferred embodiment of the present invention and other advanced methods in academia. DETAILED DESCRIPTION

[0045] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several variations and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0046] The present invention provides an embodiment of a method for combining face recognition and face image quality assessment using an adaptive mechanism, comprising:

[0047] S100, constructing an adaptive scale factor related to the feature amplitude;

[0048] S200, constructs an angle penalty and forms a joint adaptive mechanism with the scale factor in S100;

[0049] S300, constructing a regularization term that is adaptive to the feature amplitude;

[0050] S400, using the scale factor in S100, the angle penalty in S200, and the regularization term in S300 to form a cross entropy loss function in the angle space for face recognition and face image quality evaluation;

[0051] S500, constructing a face feature extractor, which extracts face features and obtains feature amplitudes;

[0052] S600, constructing a face identity determiner, which obtains a predicted identity based on facial features;

[0053] S700, based on the cross entropy loss function in S400, train the face feature extractor constructed in S500 and the face identity determiner constructed in S600, and use the trained face feature extractor and face identity determiner to complete face recognition, face identity determination and face image quality evaluation.

[0054] In this embodiment, the adaptive scaling factor links the scaling factor to the feature amplitude, enabling adaptive changes in the scaling factor based on the sample. This optimizes the distribution of features in the feature space, making the learned facial features more discriminative. This ensures the reliability and generalizability of the quality score, guaranteeing performance on recognition tasks.

[0055] In a preferred embodiment of the present invention, S100 is implemented as follows: an adaptive scale factor s(a) related to the characteristic amplitude is designed. i ), s(a i ) is about a i A linear function of i Represents facial features x i The amplitude, i.e. a i =||x i ||2. Adaptive scale factor s(a i ) is defined in [l a ,u a ] on a i The linear function of s(a i ) is in the form of:

[0056]

[0057] Among them, u a and l a A i The upper and lower bounds of u s and l s s(a i )’s upper and lower bounds.

[0058] In a preferred embodiment of the invention, S200 is implemented, specifically, to design a scale factor s(a i ) and angle penalty m(a i ) joint adaptive mechanism. Among them, the angle penalty m(a i ) is defined in [l a ,u a] on a i A monotonically increasing convex function, m(a i )'s upper and lower bounds u m and l m satisfy m(a i ) satisfies the first-order derivative of m'(a i )∈[k,K], where k satisfies

[0059] In this example, m(a i ) is in the form of:

[0060]

[0061] The joint adaptive mechanism of the scale factor and angle penalty in this embodiment means that both the scale factor and the angle penalty are functions of the feature amplitude as the independent variable, that is, they are jointly adaptive to the feature amplitude. This joint adaptive mechanism is used to optimize the distribution of the feature amplitude so that it approximately obeys the normal distribution.

[0062] In a preferred embodiment of the present invention, S300 is implemented, specifically, to design a regularization term g(a) that is adaptive to the characteristic amplitude. i ), which is defined in [l a ,u a ] on a i Strictly monotone decreasing convex function, g(a i ) satisfies the first-order derivative of g′(u a )=0.

[0063] In this example, g(a i ) is in the form of:

[0064]

[0065] The regularization term g(a i ), whose properties ensure that For about a i Convex function, and in a i The domain of definition a ,u a ]superior, There is a unique minimum point. This property makes Satisfy the basic requirements as a loss function.

[0066] In a preferred embodiment of the present invention, S400 is implemented, specifically, by combining the adaptive scaling factor s(a i ), angle penalty m(a i ) and the regularization term g(a i) for face recognition and face image quality evaluation, the cross entropy loss function in the angle space for Any one of Its specific form is:

[0067]

[0068] Among them, λ satisfies s(a i ) is the adaptive scale factor; m(a i ) is the angle penalty; g(a i ) is a regularization term; s is a fixed scale factor, and its value is u s ;

[0069] In a preferred embodiment of the present invention, S500 is implemented, the face feature extractor includes but is not limited to different structures such as multi-layer perceptron (MLP), convolutional neural network (CNN) and Transformer; the face identity determiner includes a fully connected layer; the face feature extractor and the face identity determiner adopt the loss function The face image p i The corresponding identity y i , facial features x i Amplitude a i and category distance As input, iterative optimization is performed.

[0070] In this embodiment, a loss function that adapts to the feature amplitude is used to iteratively optimize the facial feature extractor and the identity determiner, thereby obtaining more discernible facial features, more accurate identity recognition, and a quality evaluation that is more consistent with the recognition.

[0071] Furthermore, the loss function It is defined in [l a ,u a ] on a i A strictly convex function with a unique minimum point a * This property makes It meets the basic requirements of a loss function and can be used for iterative optimization of neural networks.

[0072] a * About intra-class distance Monotonically decreasing, a * About the distance between classes Monotonically increasing. For any facial image, its quality can be measured by the sample's intra-class distance and inter-class distance within the sample space. The smaller the intra-class distance, the higher the sample quality; the larger the inter-class distance, the higher the sample quality. Therefore, this property of the loss function ensures that the feature amplitude after iterative optimization can represent the quality of the facial image, resulting in a face quality score that is more consistent with recognition.

[0073] Based on the same inventive concept, an embodiment of the present invention further provides a joint system for face recognition and face image quality assessment using a feature amplitude adaptive mechanism, comprising:

[0074] The face feature extraction module is a face feature extractor trained by the combined method of face recognition and face image quality evaluation in any of the above embodiments. i , use the face feature extractor to extract its facial features x i , x i Amplitude a i That is the quality score of the face image.

[0075] The face identity determination module is a face identity determiner trained by the combined method of face recognition and face image quality assessment in any of the above embodiments. For any face feature x i , use the face identity determiner to get the corresponding predicted identity

[0076] The above-mentioned facial feature extraction module and facial identity determination module complete facial recognition, facial identity determination and facial image quality evaluation.

[0077] The modules / units in the above examples of the present invention may specifically refer to the implementation techniques of the corresponding steps of the combined method of face recognition and face image quality assessment with feature amplitude adaptive mechanism in the above embodiments, which will not be described in detail here.

[0078] Based on the same inventive concept, in other embodiments of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the terminal can be used to execute any one of the methods described, or to run the system described.

[0079] Based on the same inventive concept, in other embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, it can be used to perform any of the methods described above, or to run the system described above.

[0080] The methods provided in the examples are described in further detail below.

[0081] This embodiment consists of four steps. The first step is to construct a loss function for the feature amplitude adaptation mechanism, achieving a more normal distribution for the feature amplitudes and expressing quality scores using the feature amplitudes. This is used for iterative optimization of the facial feature extractor and facial identity detector. The specific process is similar to the implementation techniques corresponding to steps S100-S400 in the above embodiment and will not be further described here.

[0082] Step 2: Construct a neural network and implement a face feature extraction module to extract features from the input face image.

[0083] The specific operations in this step are:

[0084] 1. Design a neural network FNN as a face feature extractor FNN. For any face image p i , use the face feature extractor FNN to extract its facial features x i , x i Amplitude a i That is the quality score of the face image.

[0085] x i =FNN(p i )

[0086] a i =||x i ||

[0087] Here, ||·|| represents the 2-norm of the vector.

[0088] In this example, the structure of the FNN adopts a 100-layer convolutional neural network (CNN), and its specific parameters are shown in Table 1.

[0089] Table 1

[0090]

[0091] Step 3: Construct a neural network to realize face identity determination, which is used to determine the corresponding identity of the extracted facial features.

[0092] The specific operations in this step are:

[0093] 1. Design a neural network INN as a face feature extractor INN, for the face feature x proposed by the face feature extractor FNN i , use the face identity determiner INN to determine its corresponding predicted identity

[0094]

[0095] in, Indicates taking the identity id corresponding to the minimum value in the category distance vector.

[0096] In this example, the INN structure adopts a 1-layer fully connected layer, and its specific parameters are shown in Table 2.

[0097] Table 2

[0098] Layer Name Input size Output size FC1 1×512 1×85742

[0099] Step 4: The face feature extractor FNN and the face identity detector INN use the loss function together The face image p i The corresponding identity y i , facial features x i Amplitude a i and category distance As input, use the loss function Calculate the corresponding loss, backpropagate the loss to FNN and INN, and perform iterative optimization.

[0100] In this example, the optimizer for iterative optimization is SGD, the number of iterations is 25, and the learning rate is 0.1 from 1 to 10 rounds, 0.01 from 11 to 18 rounds, 0.001 from 18 to 22 rounds, and 0.0001 from 23 to 25 rounds.

[0101] Figure 1 This is a flowchart illustrating the entire method. It details the specific processes and data flows for steps 1 to 4. In this embodiment, the feature amplitude adaptation mechanism is integrated into a single loss function, which is used to optimize the neural network to simultaneously perform both face recognition and facial image quality assessment. This method is easily reproducible and has good applicability and generalizability.

[0102] In summary, the method steps in this embodiment design a feature amplitude adaptive mechanism. By introducing a joint adaptive mechanism of scale factor and angle penalty, the iteratively optimized feature amplitude can represent the quality of the facial image, ensuring that the distribution of the feature amplitude follows a nearly normal distribution. This reasonable feature distribution not only makes the learned facial features more discriminative and improves the accuracy of identity recognition, but also improves the consistency between quality evaluation and face recognition, effectively distinguishing high-quality and low-quality facial images. This makes this method achieve state-of-the-art results in both face recognition and face image quality assessment.

[0103] Tables 3 and 4 below show numerical comparisons of the final recognition accuracy of face recognition performance achieved using the method in this embodiment. From top to bottom, the numerical comparisons of other results with the implementation results of this embodiment (ScaleFace) are presented. It can be seen that this embodiment significantly improves face recognition accuracy across various datasets.

[0104] Table 3

[0105]

[0106] Table 4

[0107]

[0108] See also Figure 2 , which shows the comparison results of the EVRC curves for face quality assessment performance obtained in this embodiment. The closer this curve is to the origin, the more consistent the corresponding quality assessment method is with the recognition results, and the better the performance of the quality assessment method. Comparing the other result curves used for comparison with the curve of the implementation results (ScaleFace) of this embodiment, we can see that the face recognition accuracy of this embodiment is significantly improved using different recognition algorithms on different data sets.

[0109] See also Figure 3 , which is the comparison result of the face quality score distribution obtained in this embodiment. Comparing the score distribution of other comparison scores with the score distribution of this embodiment (ScaleFace), we can see that the face quality score distribution of this embodiment is closer to a normal distribution, thus ensuring consistency between quality evaluation and identity recognition.

[0110] See Table 3, Table 4, Figure 2 and Figure 3 , Figure 2 Different rows represent different data sets, and different columns represent the performance evaluation of multiple quality assessment algorithms using different face recognition algorithms; Figure 3 Each figure represents a different dataset. It can be seen that the feature amplitude-dependent scaling factor and the angle-penalty adaptive loss function proposed in this embodiment can effectively optimize the distribution of facial features in the feature space, learn more discriminative facial features, and improve face recognition accuracy. Furthermore, the feature amplitude adaptive mechanism ensures consistency between quality evaluation and face recognition, making the quality evaluation results of this method highly generalizable across different recognition methods and datasets.

[0111] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various modifications or variations within the scope of the claims without affecting the essence of the present invention. The above preferred features may be used in any combination as long as they do not conflict with each other.

Claims

1. A combined method of face recognition and face image quality assessment based on an adaptive mechanism, characterized in that: include: Constructing an adaptive scale factor related to the feature amplitude; Constructing an angle penalty and forming a joint adaptive mechanism with the scale factor; Construct a regularization term that is adaptive to the feature amplitude; Using the scale factor, angle penalty and regularization term, a cross entropy loss function in angle space is formed for face recognition and face image quality assessment; Constructing a face feature extractor to extract face features and obtain feature amplitudes; constructing a face identity determiner that obtains a predicted identity based on the face features; The face feature extractor and face identity determiner are trained based on the cross entropy loss function, and the trained face feature extractor and face identity determiner are used to complete face recognition, face identity determination and face image quality evaluation; the adaptive scale factor is s(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i The linear function of Among them, u a and l a are characteristic amplitudes a i The upper and lower bounds of u s and l s s(a i )’s upper and lower bounds; The angle penalty is m(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i A monotonically increasing convex function of , where: m(a i )'s upper and lower bounds u m and l m satisfy m(a i ) satisfies the first-order derivative of m′(a i )∈[k,K], k satisfies K is the upper bound of the first-order derivative of the angle penalty m'(a); The regularization term is g(a i ), is defined in [l a ,u a ] on the characteristic amplitude a i Strictly monotone decreasing convex function, g(a i ) satisfies the first-order derivative of g ′ (u a )=0,u a and l a A i The upper and lower bounds of .

2. The combined method of face recognition and face image quality assessment based on an adaptive mechanism according to claim 1, characterized in that: Using the scale factor, angle penalty and regularization term, the cross entropy loss function in the angle space for face recognition and face image quality evaluation is formed as follows: The cross entropy loss function Any one of That is, for a single face image p i The loss function is: Among them, λ satisfies s(a i ) is the adaptive scale factor; s is the fixed scale factor, which takes the value of u s ;m(a i ) is the angle penalty; g(a i ) is a regular term; N represents the batch size, i represents the i-th face sample in the batch, x i Represents facial features, y i Represents x i Belongs to the identity category, θ j represents the j-th identity category weight W j with x i The angle between them, that is, the category distance, where j = 1, 2, 3, ..., C (j ≠ y i ), C is the total number of identities.

3. The combined method of face recognition and face image quality assessment based on an adaptive mechanism according to claim 2, characterized in that: Design a neural network FNN as a face feature extractor FNN; Design a neural network INN as a face identity determiner INN; The face feature extractor FNN and the face identity determiner INN use the loss function to convert the identity corresponding to the face image, the amplitude of the face feature and the category distance As input, iterative optimization is performed.

4. The combined method of face recognition and face image quality assessment based on an adaptive mechanism according to claim 2, characterized in that: The loss function It is defined in [l a ,u a ] on a i A strictly convex function with a unique minimum point a * ;a * About intra-class distance Monotonically decreasing, a * About the distance between classes Monotonically increasing.

5. A joint system of face recognition and face image quality assessment with an adaptive mechanism, characterized by: include: A facial feature extraction module, comprising a facial feature extractor trained using the method of any one of claims 1 to 4. For any facial image, the facial feature extractor extracts facial features, and the magnitude of the facial features is the quality score of the facial image; A face identity determination module, comprising a face identity determiner trained using the method of any one of claims 1 to 4, wherein for any facial feature, the face identity determiner obtains the corresponding predicted identity; The face feature extraction module and the face identity determination module complete face recognition, face identity determination and face image quality evaluation.

6. A combined terminal for face recognition and face image quality assessment, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it can be used to perform the method according to any one of claims 1 to 4, or run the system according to claim 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it can be used to perform the method according to any one of claims 1 to 4, or to run the system according to claim 5.

Images