Method and system for face image quality evaluation based on learning ranking
By constructing a standardized intra-class and inter-class distance based on learning-based ranking for face image quality assessment, and combining it with a visual Transformer network, the face image quality assessment is optimized, solving the bias problem in existing technologies and improving recognition accuracy and applicability.
Patent Information
- Application Number
- CN202211540806.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-02
- Publication Date
- 2025-12-19
- Estimated Expiration
- 2042-12-02
AI Technical Summary
Existing face image quality assessment methods have biases, leading to a decrease in recognition accuracy. In particular, low-quality images are prone to misidentification in practical applications, and the backbone network is overfitted to a specific recognition algorithm, making it difficult to adapt to multiple recognition systems.
A face image quality assessment method based on learning-based ranking constructs class centers for each identity in a face dataset, calculates standardized intra-class and inter-class distances, generates quality labels, and optimizes the quality assessment by training a visual Transformer network and a cross-entropy loss function.
It reduces bias factors in existing methods, improves the performance of quality assessment and the consistency of identification results, has wide applicability, and is easy to reproduce and promote.
Smart Images
Figure CN115731599B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer vision, in particular, to a face image quality evaluation method and system based on learning ranking, a medium and a terminal. BACKGROUND
[0002] Face image contains a lot of information about personal gender, age, skin color, appearance characteristics, etc., and is an important personal information that can be used for biometric recognition. With the progress of computer vision and deep learning and other related technologies, the recognition accuracy of computers for the identity corresponding to the face image has been greatly improved, and various face recognition systems have gradually begun to popularize in daily life and play a role in identity recognition and authentication in various aspects of society. However, face recognition systems are often greatly challenged in actual application. Compared with the face images in the data set which are standardized and aligned, the face images in the actual application scene are of uneven quality due to the constraints of the collection conditions. The accuracy of the identity recognition of the face recognition system is often closely related to the quality of the input face image, and low-quality face images are extremely easy to cause misrecognition, thereby causing security risks. Therefore, in actual application, the face recognition system is faced with great challenges from low-quality face images and is difficult to achieve the high accuracy in academic experiments. In order to solve this problem, the academic field proposes to add a face image quality evaluation algorithm before the recognition link to eliminate low-quality face images, thereby improving the reliability of recognition.
[0003] From the existing technical literature retrieval, it is found that there are two main research ideas for face image quality evaluation methods for machine recognition. One idea is to train a regression model based on identity labels. This kind of method analyzes the essence of face image quality-the intra-class distance and inter-class distance of face features in the latent space, and generates the quality label of the face image according to the distance, and uses the label to train the regression model to evaluate the face quality. (See: Hernandez-Ortega, Javier, 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; Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, 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; Boutros, Fadi, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. CR-FIQA: face image quality assessment by learning sample relative classifiability. ArXiv preprint arXiv:2112.06592 (2021).) However, the quality label in the above method is based on the absolute distance, which will cause deviation from the true quality due to uneven sample data distribution, and the use of regression model will further amplify this deviation.
[0004] Another idea is to implement a face recognition and quality evaluation joint method, which provides a solution to optimize the distribution of face features in 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 learning of face quality of these methods is often based on implicit constraints, which is difficult to reach the optimal point in actual training and thus deviates from the real quality. In addition, the backbone network of all the above methods is related to a specific face recognition algorithm, which will also make the face quality evaluation deviate too much from the specific recognition algorithm and produce a large deviation from other recognition algorithms. SUMMARY
[0005] In view of the defects in the prior art, the purpose of the present application is to provide a face image quality evaluation method and system based on learning ranking.
[0006] According to one aspect of the present application, a face image quality evaluation method based on learning ranking is provided, comprising:
[0007] extracting face features of the face image in the latent space using a face recognition model, and constructing a class center corresponding to each identity in the face data set based on the face features;
[0008] constructing a standardized intra-class distance of the face image sample in the latent space based on the class center;
[0009] constructing a standardized inter-class distance of the face image sample in the latent space based on the class center;
[0010] constructing a quality label of the face image sample using the standardized intra-class distance and the standardized inter-class distance of the face image sample;
[0011] Constructing a face image quality network for predicting quality of face images;
[0012] Constructing pairs of face images According to their quality labels, constructing with a higher quality than the actual probability as training labels; according to the quality network predictions of their quality, constructing with a higher quality than the posterior probability;
[0013] According to the actual probability and the posterior probability of the pairs of face images, constructing a cross-entropy loss function;
[0014] Training the face image quality network based on the cross-entropy loss function, and using the trained face image quality network to complete face image quality evaluation.
[0015] Preferably, the face recognition model is used to extract face features of the face image in the latent space, and a class center corresponding to each identity in the face data set is constructed based on the face features, including:
[0016] Given a face image data set containing m identities where represents the i-th face image of the j-th identity, n j represents the total number of images of the j-th identity;
[0017] For Using a pre-trained face recognition model to extract the corresponding face feature vector:
[0018] For any identity j, the class center is constructed using the features of all face images of the identity:
[0019]
[0020] Preferably, the class center is used to construct a standardized intra-class distance of the face image sample in the latent space, including:
[0021] For any face image calculate the intra-class angular distance to the corresponding class center:
[0022] Using the intra-class angular distances of all face images in the same identity, calculate the mean value of the intra-class distance of the identity: and the standard deviation:
[0023] The standardized inter-class distance of the face image sample in the latent space is constructed based on the intra-class angle distance, the intra-class distance average value and the standard deviation of the same identity class:
[0024]
[0025] Preferably, the standardized inter-class distance of the face image sample in the latent space is constructed based on the class center, comprising:
[0026] For any face image The angle distance from the center of the identity class k is calculated: k = 1, 2, …, m, k≠j;
[0027] The smallest angle distance to other classes is selected as the inter-class distance of the identity class:
[0028] The average value of the intra-class distance of the identity class with the smallest angle distance is calculated And the standard deviation
[0029] The standardized inter-class distance of the face image sample in the latent space is constructed based on the average value and the standard deviation of the intra-class distance of the smallest class:
[0030]
[0031] Preferably, the quality label of the face image sample is constructed based on the standardized intra-class distance and the standardized inter-class distance of the face image sample:
[0032]
[0033] Where β is a constant.
[0034] Preferably, the face image quality network is a visual Transformer-based neural network, and the input is a face image, and the output is a quality score of the face image.
[0035] Preferably, the face image pair The quality label of the face image pair is The actual probability that the quality is higher than is used as the training label:
[0036]
[0037] The face image quality network is used to predict the quality score of the face image respectively: and constructing the higher than posterior probability:
[0038]
[0039] Preferably, according to the actual probability of the face image pair and the posterior probability P 12 , a cross-entropy loss function is constructed:
[0040]
[0041] The face image quality network and the face image pair are substituted, and the cross-entropy loss function is constructed:
[0042]
[0043] The face image quality network adopts the loss function, takes the face image pair and the training label as input, and performs iterative optimization.
[0044] According to a second aspect of the present application, a system for face image quality evaluation based on learning ranking is provided, comprising:
[0045] A label generation module extracts face features of face images in a latent space by using a face recognition model, and constructs a class center corresponding to each identity in a face data set based on the face features;
[0046] Based on the class center, a standardized intra-class distance of face image samples in the latent space is constructed;
[0047] Based on the class center, a standardized inter-class distance of face image samples in the latent space is constructed;
[0048] The standardized intra-class distance and the standardized inter-class distance of the face image samples are used to construct a quality label of the face image samples;
[0049] A face image quality evaluation module constructs a face image quality network for predicting the quality of face images;
[0050] A learning ranking module constructs a face image pair according to their quality labels, and constructs higher than actual probabilities as training labels; constructing a cross-entropy loss function according to the actual probabilities and the posterior probabilities of the face image pairs; the posterior probabilities of the face image pairs; the posterior probabilities of the face image pairs;
[0051] constructing a cross-entropy loss function according to the actual probabilities and the posterior probabilities of the face image pairs;
[0052] training the face image quality network based on the cross-entropy loss function, and completing face image quality evaluation by using the trained face image quality network.
[0053] According to a third aspect of the present application, there is provided a terminal for face image quality evaluation based on learning ranking, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the program to perform any of the above methods or run the above system.
[0054] According to a fourth aspect of the present application, there is provided a computer readable storage medium having a computer program stored thereon, wherein the program is executable by a processor to perform any of the above methods or run the above system.
[0055] Compared with the prior art, the present application has at least one of the following beneficial effects:
[0056] 1. The face image quality evaluation method and system based on learning ranking in the embodiment of the present application, based on the face image quality labels of standardized intra-class distance and inter-class distance, the quality network based on visual Transformer and the training method based on pair-wise learning ranking, are used to reduce the bias factors existing in the existing face image quality evaluation method, thereby improving the performance of quality evaluation and obtaining face image quality scores more consistent with the recognition results. It is easy to reproduce and has good applicability and generalizability.
[0057] 2. The face image quality evaluation method and system based on learning ranking in the embodiment of the present application, based on the intra-class angle distance between the same identity samples in the latent space and the inter-class angle distance between the different identity samples, the quality labels of each sample are generated, which solves the problem that the distribution of different identity face images in age, posture, shape and other factors is different, resulting in inconsistency between the quality labels based on absolute distance and the real quality.
[0058] 3. The face image quality evaluation method and system based on learning ranking in the embodiment of the present application, a face image quality evaluation model based on visual Transformer is designed, which improves the applicability of the backbone network to the quality evaluation task and eliminates the problem of overfitting to the feature recognition method.
[0059] 4、The method and system for face image quality evaluation based on learning ranking in the embodiment of the application adopt a learning ranking algorithm framework of pair training to train a quality model, which can effectively solve the deviation caused by direct use of a regression model fitting and improve the consistency of quality evaluation and recognition. BRIEF DESCRIPTION OF DRAWINGS
[0060] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments with reference to the drawings:
[0061] Figure 1 A flowchart of the method for face image quality evaluation based on learning ranking in a preferred embodiment of the application is shown in the figure.
[0062] Figure 2 A structural diagram of the method for face image quality evaluation based on learning ranking in a preferred embodiment of the application is shown in the figure.
[0063] Figure 3 A structural diagram of the backbone network based on visual Transformer in a preferred embodiment of the application is shown in the figure.
[0064] Figure 4 A structural diagram of the SwinTransformer block in the backbone network based on visual Transformer in a preferred embodiment of the application is shown in the figure.
[0065] Figure 5 A comparison result curve of face image quality evaluation and other advanced methods in the academic field in a preferred embodiment of the application is shown in the figure. DETAILED DESCRIPTION
[0066] The application 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 application, but do not limit the application in any form. It should be noted that, for those skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made. These all belong to the protection scope of the application.
[0067] Reference Figure 1 and Figure 2 The application provides an embodiment of a method for face image quality evaluation based on learning ranking, which comprises:
[0068] S100, extracting face features of a face image in a latent space by using a face recognition model, and constructing a class center corresponding to each identity in a face data set based on the face features;
[0069] S200, constructing a standardized intra-class distance of a face image sample in the latent space based on the class center constructed in S100;
[0070] S300, constructing a standardized inter-class distance of the face image samples in the latent space based on the class center constructed in S100;
[0071] S400, constructing a quality label of the face image samples by using the standardized intra-class distance and the standardized inter-class distance of the face image samples in S200 and S300;
[0072] S500, constructing a face image quality network for predicting the quality of the face image;
[0073] S600, constructing face image pairs, constructing a prior probability that the quality is higher than the threshold as a training label according to the quality label of the face image pairs constructed in S100-S400, and constructing a posterior probability that the quality is higher than the threshold according to the prediction of the quality of the face image pairs by the quality network in S500;
[0074] S700, constructing a cross-entropy loss function according to the prior probability and the posterior probability of the face image pairs in S600;
[0075] S800, training the face image quality network based on the cross-entropy loss function in S700, and completing the face image quality evaluation by using the trained face image quality network.
[0076] In the embodiment, the face image quality label based on the standardized intra-class distance and the inter-class distance, the quality network based on the visual Transformer, and the training method based on the pair-wise learning ranking are used to reduce the bias factors existing in the existing face image quality evaluation method, so as to improve the performance of the quality evaluation, and obtain the face image quality score more consistent with the recognition result. It is easy to reproduce and has good applicability and generalization.
[0077] In a preferred embodiment of the present application, S100 is implemented, and the specific steps are as follows:
[0078] The face features of the face image in the latent space are extracted by using the face recognition model, and the class center corresponding to each identity in the face data set is constructed, specifically, given a face image data set containing m identities Wherein represents the i-th face image of the j-th identity, n j represents the total number of images of the j-th identity. For A pre-trained face recognition model is used The corresponding face feature vector is extracted:
[0079] For any identity j, the class center is constructed by using the features of all face images of the identity:
[0080]
[0081] In a preferred embodiment of the present application, S200 is implemented to construct the normalized intra-class distance of the face image sample in the latent space using the class center in S100, specifically, for any face image calculate the intra-class angle distance to the corresponding class center:
[0082]
[0083] Using the intra-class angle distance of all face images in the same identity, the average value of the intra-class distance of the identity is calculated: and the standard deviation: Using the intra-class distance, the average value and the standard deviation of the intra-class distance of the identity, the normalized intra-class distance of the face image sample in the latent space is constructed:
[0084]
[0085] In a preferred embodiment of the present application, S300 is implemented to construct the normalized inter-class distance of the face image sample in the latent space using the class center in S200, specifically, for any face image calculate the angle distance to the center of other identity class k (k = 1, 2, …, m, k ≠ j):
[0086]
[0087] Select the smallest one among all the angle distances to other classes as the inter-class distance of : According to the average value of the intra-class distance of the minimum class and the standard deviation construct the normalized inter-class distance of the face image sample in the latent space:
[0088]
[0089] In a preferred embodiment of the present application, S400 is implemented to construct the quality label of the face image sample using the normalized intra-class distance of the face image sample in S200 and the inter-class distance in S300, specifically, using the normalized intra-class distance of the face image sample and the normalized inter-class distance construct the quality label of the face image sample:
[0090]
[0091] wherein β is a constant, and in this embodiment, 0.1 is taken.
[0092] The above embodiment adopts the standardized intra-class angle distance and inter-class angle distance of the face features in the latent space to generate the quality label of each sample, solves the problem that the absolute distance quality label does not match the real quality due to the distribution difference of face images of different identities in terms of age, posture, style and other factors, improves the accuracy of the face quality training label, and thus improves the performance of the quality evaluation.
[0093] In a preferred embodiment of the present application, S500 is implemented to construct a face image quality network for predicting the quality score of an input face image. The specific steps are as follows:
[0094] Constructing a face image quality network based on visual Transformer In this embodiment, the specific structure of the network is as shown in Figure 3 An input face image with a size of 112x112 is input, a patch partition layer is used to deform and increase the dimension of the input image to achieve the required input shape. Then the feature continuously passes through three Swin layers, each Swin layer contains multiple Swin Transformer blocks, and a patch merging layer, and the last Swin layer does not contain a patch merging layer, but an average pooling layer. The function of the patch merging layer is to deform and reduce the dimension of the feature. After passing through the three Swin layers, a one-dimensional vector is finally output, and after passing through a fully connected layer, the quality score can be output. The specific parameters of each Swin layer are shown in Table 1.
[0095] Table 1 Specific parameters of the face image quality network in this embodiment
[0096] Swin layer Input size Window size Depth Number of attention heads Swin-Layer 1 28×28 7 2 3 Swin-Layer 2 14×14 7 18 12 Swin-Layer 3 7×7 7 2 12
[0097] The structure of the Swin Transformer block in this embodiment is as shown in Figure 4 As shown, it mainly consists of a Window Multihead Self Attention (W-MSA) layer, a Shifted Window Multihead Self Attention (SW-MSA) layer, a Multi-Layer Perceptron (MLP), and a Layer Normalization (LN) layer. SW-MSA and W-MSA alternate in the model; that is, if the attention layer in the previous Swin-Transformer block is W-MSA, then the attention layer in the next Swin-Transformer block is SW-MSA.
[0098] In a preferred embodiment of the present invention, step S600 is implemented to construct face image pairs, and their quality labels are constructed according to steps S100-S400, with the actual probability of higher quality being used as the training label; based on the quality prediction of their quality by the quality network in step S500, the posterior probability of higher quality is constructed. The specific steps are as follows:
[0099] Construct face image pairs Based on their quality labels, construct Quality higher than The actual probabilities are used as training labels; based on the quality network's predictions of their quality, a model is constructed. Quality higher than The posterior probability, specifically, is the construction of face image pairs. Its quality label is The Quality higher than The actual probability is used as the training label:
[0100]
[0101] Using the face image quality network Predict face images separately and mass fraction:
[0102]
[0103] Construct the above using the Sigmoid function Quality higher than Posterior probability:
[0104]
[0105] In a preferred embodiment of the present application, S700 is implemented, and the actual probability and the posterior probability of the face image pair in S600 are used to construct a cross-entropy loss function, specifically,
[0106]
[0107] The face image quality network is substituted, and a cross-entropy loss function for training the face image quality network is constructed:
[0108]
[0109] The face image quality network is trained based on the above cross-entropy loss function, specifically, for each face image x1 in the data set, an arbitrary other face image x2 in the data set is selected to pair with it to become a random face image pair. The actual probability and the posterior probability that x1 is of higher quality than x2 are constructed based on the method in S31, and then the loss is calculated using the cross-entropy loss function, and the parameters are updated in the face image quality network In this example, the optimizer of the iterative optimization is SGD, the number of iterations is 60, the learning rate is 0.1 for 1-15 rounds, 0.01 for 16-30 rounds, 0.001 for 31-45 rounds, and 0.0001 for 46-60 rounds.
[0110] In another embodiment of the present application, the face image quality network trained in S700 is used to evaluate the quality of a face image, specifically, a single face image is input into the face image quality network , and the quality score corresponding to the image can be obtained.
[0111] In the above embodiments, the quality label of each sample is generated by using the standardized intra-class and inter-class distances of the face features in the latent space, the backbone network based on visual Transformer is introduced, and the training is performed by using the method framework of pair-wise learning ranking, which effectively eliminates the factors that are prone to cause evaluation bias in existing face image quality evaluation methods, enhances the consistency of face quality evaluation and recognition, and improves the performance of face image quality evaluation.
[0112] Based on the same inventive concept, the embodiments of the present application also provide a face image quality evaluation system based on learning ranking, which includes a quality label generation module, a face image quality evaluation module, and a learning ranking module; wherein,
[0113] a quality label generation module configured to extract facial features of the face image in a latent space by using the face recognition model, construct a class center corresponding to each identity in the face data set based on the facial features, construct a standardized intra-class distance of the face image sample in the latent space based on the class center, construct a standardized inter-class distance of the face image sample in the latent space based on the class center, and construct a quality label of the face image sample by using the standardized intra-class distance and the standardized inter-class distance of the face image sample;
[0114] a face image quality evaluation module configured to construct a face image quality network for predicting the quality of the face image;
[0115] a learning ranking module configured to construct a face image pair according to the quality labels thereof, construct a posterior probability that the quality is higher than as a training label according to a prediction of the quality of the face image pair by using the face image quality network , and construct a posterior probability that the quality is higher than according to the actual probability and the posterior probability of the face image pair, construct a cross-entropy loss function according to the actual probability and the posterior probability of the face image pair, and train the face image quality network based on the cross-entropy loss function, so as to complete face image quality evaluation by using the trained face image quality network.
[0116] In the embodiment, a quality label generation module based on standardized intra-class and inter-class distances is designed to generate a quality label with less deviation for the face image, a backbone network based on visual Transformer is introduced to improve the basic adaptability of the network to the task, and a paired learning ranking algorithm framework is used to train the quality network to eliminate the deviation caused by directly using a regression algorithm. The modules in the embodiment eliminate factors that are prone to cause evaluation deviation in existing face image quality evaluation methods from multiple aspects, enhance the consistency of face quality evaluation and recognition, improve the performance of face image quality evaluation, and reach the most advanced level in the academic field.
[0117] In the above examples of the present application, each module can refer to the implementation techniques of the corresponding steps of the face image quality evaluation method based on learning ranking in the above embodiments, which will not be described here.
[0118] Based on the same inventive concept, in other embodiments of the present application, a terminal is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor can be used to execute any of the methods described above, or run the system described above.
[0119] Based on the same inventive concept, in other embodiments of the present application, a computer readable storage medium is provided, which stores a computer program, the program being executable by a processor to implement any of the methods described above, or to operate the system described above.
[0120] Referring to Figure 5 is the comparison result of the EVRC curve of the face quality evaluation performance obtained by the embodiment. The closer the curve is to the origin, the more consistent the corresponding quality evaluation method is with the recognition result, and the better the quality evaluation method performance is. From the comparison of the curve of the embodiment (L2RT-FIQA) with the curves of other results used for comparison, it can be seen that the face image quality evaluation performance of the embodiment has good effect improvement on different recognition algorithms and different data sets.
[0121] Referring to Table 2, which is the AOC result of the face quality evaluation performance obtained by the embodiment and the comparison of other advanced algorithms in the academic field, the larger the AOC value, the better the face image quality evaluation performance. From the comparison of the results of the embodiment (L2RT-FIQA) with other results used for comparison, the performance of the embodiment is optimal in 21 groups of 27 experiments, and it can be seen that the embodiment can achieve a very high level under different false match rates, which reflects a very high generalization ability.
[0122] Table 2
[0123]
[0124] In summary, the face image quality evaluation result of the method shows very high quality evaluation performance for different recognition methods, different data sets, and different false rejection rates, and has good generalization ability, which proves the effectiveness and rationality of the method.
[0125] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various modifications or changes within the scope of the claims, which does not affect the essential content of the present application. The above preferred features can be used in combination in the case of not conflicting with each other.
Claims
1. A method for face image quality evaluation based on learning ranking, characterized in that, The method comprises the following steps: extracting facial features of a face image in latent space by using a face recognition model, and constructing a class center corresponding to each identity in a face data set based on the facial features; constructing a standardized intra-class distance of a face image sample in latent space based on the class center; constructing a standardized inter-class distance of a face image sample in latent space based on the class center; constructing a quality label of a face image sample by using the standardized intra-class distance and the standardized inter-class distance of the face image sample; constructing a face image quality network Q for predicting the quality of a face image; Constructing pairs of face images Constructing pairs of face images with a quality higher than the actual probability as training labels; constructing pairs of face images with a posterior probability higher than the actual probability as training labels; constructing a cross-entropy loss function according to the actual probability and the posterior probability of the face image pair; training the face image quality network based on the cross-entropy loss function, and completing face image quality evaluation by using the trained face image quality network; The face image quality network Q is a neural network based on a visual Transformer, and the input is a face image and the output is a quality score of the face image. said pair of face images whose quality label is said higher than the actual probability as training labels: Using the face image quality network Q, a quality score of a face image and is predicted respectively: The posterior probability that the face image has a quality higher than is constructed: According to the actual probability of the face image pair and the posterior probability P 12 , a cross-entropy loss function is constructed: The face image quality network and face image pairs Substitute, construct the cross-entropy loss function: The face image quality network With the loss function, the face image pair And the training label As input, iterative optimization is performed. 2.The method of claim 1, wherein, The method comprises the following steps: Given a face image dataset containing m identities wherein denotes the i-th face image of the j-th identity, n j denotes the total number of images of the j-th identity; For extracting a corresponding face feature vector using a pre-trained face recognition model R: For any identity j, the class center is constructed by using the features of all face images of the identity: 3.The method of claim 1, wherein, The standardized intra-class distance of the face image sample in latent space is constructed by using the intra-class angle distance, the intra-class distance average value and the standard deviation of the same identity: For any face image Compute its intra-class angular distance to the corresponding class center: The average value of the intra-class distance of the same identity is calculated by using the intra-class angle distance of all face images in the same identity: and the standard deviation: The standardized inter-class distance of the face image sample in latent space is constructed according to the intra-class distance average value and the standard deviation of the smallest class: The standardized intra-class distance and the standardized inter-class distance of the face image sample are used to construct the quality label of the face image sample: For any face image Compute its angle distance to the center of other identity class k: k = 1, 2, …, m, k≠j; The smallest of all the angle distances to other classes is chosen as the interclass distance of : calculating the average of the intra-class distances of the identities with the smallest angular distance and the standard deviation Wherein, β is a constant.
4. The method of claim 3, wherein, The method comprises the following steps: A quality label generation module extracts facial features of a face image in latent space by using a face recognition model, and constructs a class center corresponding to each identity in a face data set based on the facial features; 5. A system for learning-based ranking of face image quality assessment, characterized in that, constructing a standardized intra-class distance of a face image sample in latent space based on the class center; constructing a standardized inter-class distance of a face image sample in latent space based on the class center; constructing a quality label of a face image sample by using the standardized intra-class distance and the standardized inter-class distance of the face image sample; constructing a cross-entropy loss function according to the actual probability and the posterior probability of the face image pair; training the face image quality network based on the cross-entropy loss function, and completing face image quality evaluation by using the trained face image quality network; A face image quality evaluation module, which constructs a face image quality network for predicting the quality of a face image; a learning ranking module that constructs pairs of face images according to their quality labels a posteriori probability that the quality is higher than the actual probability as training labels; according to the quality network predictions of their quality, a posteriori probability that the quality is higher than The processor executes the program to execute the method of any one of claims 1-4, or run the system of claim 5. The program is executed by the processor to execute the method of any one of claims 1-4, or run the system of claim 5. The face image quality network is a neural network based on visual Transformer, input is a face image, and output is a quality score of the face image; said pair of face images whose quality label is said higher than the actual probability as training labels: using the face image quality network respectively predicting a quality score for a face image and constructing the posterior probability that the quality is higher than According to the face image pair The actual probability And the posterior probability P 12 , the cross-entropy loss function is constructed: The face image quality network and the face image pair Substitute, construct the cross-entropy loss function: The face image quality network With the loss function, the face image pair And the training label As input, iterative optimization is performed.
6. A terminal comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, 7. A computer-readable storage medium having stored thereon a computer program, characterized in that