Face recognition method and device based on deep semantic perception

Abstract

The invention belongs to the technical field of computer vision and pattern recognition, and relates to a face recognition method and device based on deep semantic perception, and the method comprises the steps: obtaining a face image training set, and constructing a visual feature matrix of the face image training set; performing singular value decomposition on the matrix to obtain a first-order standard projection matrix; introducing a learnable matrix to construct a second-order projection matrix, constructing a two-stage projection reconstruction matrix, and introducing a pre-trained face representation network to construct a loss function; iteratively solving an optimal learnable matrix according to the loss function, and obtaining an optimal second-order projection matrix according to the matrix; obtaining a to-be-recognized face image, and recognizing the to-be-recognized face image according to the matrix and the face image training set to obtain a face recognition result; according to the method, the sample is projected into the high-level semantic space of the network for subspace learning, so that the discrimination stability and interpretability under complex disturbance are improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and pattern recognition technology, and relates to a face recognition method and device based on deep semantic perception. Background Technology

[0002] Principal Component Analysis (PCA), a classic unsupervised linear dimensionality reduction method, has long been used for the compression and representation of high-dimensional visual data. Its statistical essence is to perform eigenvalue decomposition on the sample covariance matrix to capture the direction of maximum variance, geometrically equivalent to finding the projection hyperplane in high-dimensional space that best preserves the data structure. This method is renowned for its theoretical clarity and computational efficiency, and is widely applied in image processing, background modeling, spectral recognition, and target tracking. However, traditional PCA is based on the squared L2 norm, making it highly sensitive to noise and outliers. Extreme samples can be amplified, distorting the estimated subspace model, a problem particularly prominent in visual tasks. To improve robustness, early research attempted to replace the squared error with statistics such as absolute bias or median. Subsequently, variants such as L1-PCA and R1-PCA emerged, using L1 or L2,1 norms on the reconstructed residuals to reduce the impact of outliers. However, these variants also introduce problems such as insufficient rotation invariance, poor interpretability, or local minima caused by non-convex optimization. Related reviews and experiments show that even with improved loss, modeling based solely on the original observation space is still difficult to stably cope with complex noise distributions and structured interference.

[0003] Beyond loss function substitution, another important approach is to mitigate the contribution of unreliable samples through sample reweighting or truncation optimization. Typical examples include RPCA-OM, which jointly optimizes the sample mean to suppress noise propagation from "misaligned centering"; T-RPCA, which iteratively removes high-error samples; and RPCA-AN, which assigns continuous confidence weights to each sample to maintain local consistency. These methods can be uniformly formulated as weighted reconstruction objectives, where the weights take values ​​between binary truncation and continuous adaptation. While these strategies suppress severe outliers to some extent, they remain limited to the original pixel (or shallow, low-level feature) space, and weight design often relies on heuristic rules, resulting in insufficient ability to distinguish complex semantic structures and feature-level noise. Furthermore, subsequent research has extended robust PCA to tensor representations or introduced statistical divergence measures, achieving stronger theoretical robustness and practical effectiveness in high-dimensional or heavy-tailed distributions; however, these extensions also primarily occur at the observation space level.

[0004] In terms of the "closest existing technology," Normalized Robust PCA with Adaptive Reconstruction Error Minimization (N-RPCA) is representative. This method adaptively balances fitting and robustness at the sample level through non-convex loss and error normalization, showing strong resistance to interference, especially for noisy, occluded, or heterogeneously distributed datasets. Internal literature review also lists N-RPCA as a closely related comparison, pointing out that while it improves tolerance to anomalous samples, its adaptability to local feature noise and complex semantic structures remains insufficient. In other words, N-RPCA and similar weight / truncation paradigms focus on "sample-level" reliability modeling, making it difficult to distinguish between "feature-level" noise and identity-related discriminative information with fine granularity without compromising important semantics.

[0005] For semantically sensitive scenarios such as face recognition, recent practices have shown that modeling directly in the deep embedding space can naturally suppress some low-level interference and better preserve the identity discrimination structure. For example, using a pre-trained face representation network (such as FaceNet) to extract a 512-dimensional perceptual embedding, followed by dimensionality reduction and metric learning, can achieve lower dimensionality normalization reconstruction error on a clean set. However, if traditional projection and loss based on observation space assumptions are still used, the robustness improvement under conditions such as occlusion and drastic changes in illumination is still limited, especially when the reconstruction residual metric does not match the high-level semantic consistency. This phenomenon is clearly presented in comparative experiments and discussions: when 20% of the test samples are randomly occluded by 0.05–0.15, the recognition rate of classic PCA and various RPCA variants drops significantly, exposing the structural defect of "reconstruction-dominated but semantically mismatched".

[0006] Existing literature also points out the implicit risks of traditional methods in the "centralization" stage: decentralization using the global mean can spread the bias of contaminated samples to all samples, leading to an overall shift in subspace estimation. While methods like RPCA-OM alleviate this problem by introducing a learnable mean, a trade-off between robustness and complexity is required. Furthermore, angular objectives (such as Angle-PCA) attempt to balance fidelity and noise resistance with the ratio of "projection variance / reconstruction error," but introduce more complex non-convex optimization forms, making it difficult to guarantee the stability and global optimality of the actual solution. Overall, while alternative norms, sample weighting / truncation, joint centralization, and tensor quantization can alleviate certain types of noise or structural perturbations, they all face two fundamental constraints: first, the modeling domain often remains at the pixel or shallow feature space, resulting in inconsistencies between residuals and semantics; second, robust strategies mostly remain at the "sample level," lacking explicit measurement and constraints on the semantic sensitivity at the "feature level."

[0007] Therefore, regarding the technical problem to be solved by this invention, the existing technology has the following directly related shortcomings: First, optimal reconstruction based on the observation space is not equivalent to consistent optimality in the semantic space, which easily leads to semantic mismatch where "the reconstruction looks correct but the identity discrimination is degraded"; Second, sample-level weights or truncation may mistakenly kill effective information or retain irrelevant noise under "feature-level" damage such as occlusion and local artifacts, resulting in a decrease in the discriminability of low-dimensional representations; Third, estimation bias in centering and subspace solving will be systematically amplified in noisy environments, affecting the stability of subsequent metric learning and nearest neighbor discrimination; Finally, existing deep feature direct application schemes usually lack task-adaptive secondary projection and perceptual consistency constraints, making it difficult to obtain both low reconstruction error and high recognition robustness without sacrificing interpretability. Summary of the Invention

[0008] To address the aforementioned problems in the prior art, this invention employs a face recognition method based on deep semantic perception, comprising:

[0009] S1. Obtain the face image training set and construct the visual feature matrix of the face image training set. ;

[0010] S2, Visual feature matrix of the face image training set. Perform singular value decomposition to obtain the first-order standard projection matrix. ;

[0011] S3. Introducing a learnable matrix Constructing a second-order projection matrix Construct a two-stage projection reconstruction matrix Obtain a pre-trained face representation network Based on pre-trained face representation network Based on the visual feature matrix and two-stage projection reconstruction matrix Constructing the loss function ;

[0012] S4. Based on the loss function Iterative solution to find the optimal learnable matrix Based on the optimal learnable matrix Obtain the optimal second-order projection matrix ;

[0013] S5. Obtain the face image to be identified, and apply the optimal second-order projection matrix. Visual feature matrix of the face image training set The face image to be identified is then processed to obtain the face recognition result.

[0014] On the other hand, the present invention employs a face recognition system based on deep semantic perception to implement the above-mentioned face recognition method, including:

[0015] The visual feature matrix construction module is used to obtain a training set of face images and construct the visual feature matrix of the training set of face images. ;

[0016] The orthogonal basis construction module is used to construct the visual feature matrix of the face image training set. Perform singular value decomposition to obtain the first-order standard projection matrix. ;

[0017] The second projection and reconstruction module is used to introduce learnable matrices. Constructing a second-order projection matrix Construct a two-stage projection reconstruction matrix ;

[0018] The perceptual consistency optimization module is used to obtain the pre-trained face representation network. Based on pre-trained face representation network Based on the visual feature matrix and two-stage projection reconstruction matrix Constructing the loss function According to the loss function Iterative solution to find the optimal learnable matrix Based on the optimal learnable matrix Obtain the optimal second-order projection matrix ;

[0019] The recognition output module is used to identify the optimal second-order projection matrix. Visual feature matrix of the face image training set The face image to be identified is then processed to obtain the face recognition result.

[0020] Beneficial effects:

[0021] 1. This invention no longer minimizes reconstruction error in the original pixels or shallow feature domains, but projects samples into the high-level semantic space of a pre-trained neural network and performs subspace learning within this space. Because this semantic embedding has inherent noise suppression and semantic preservation characteristics, it can effectively suppress sample-level interference while maintaining the high-level discriminative structure closely related to recognition, thus avoiding the problem of "optimal reconstruction but semantic mismatch." This mechanism improves the discriminative stability and interpretability under complex perturbations such as noise, occlusion, and lighting changes. 2. This invention introduces a learnable quadratic projection matrix based on the orthogonal basis (i.e., the first-order standard projection matrix), forming a controllable quadratic dimensionality reduction mechanism to further refine the projection and enhance interpretability. This design allows adaptive adjustment of the dimensionality reduction intensity in the latent space, alleviating overfitting caused by excessive dimensionality truncation and enhancing the expressiveness and stability of the resulting subspace. Since the quadratic projection is optimized only within a limited subspace, it can reduce the tendency to overfit noisy samples while maintaining geometric interpretability, exhibiting more stable generalization performance across datasets. Attached Figure Description

[0022] Figure 1 A flowchart of a face recognition method based on deep semantic awareness provided in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram illustrating the principle of the deep semantic perception-based face recognition method provided in this embodiment of the invention, compared with other methods;

[0024] Figure 3 A comparison chart showing the reconstruction errors of different methods provided in the embodiments of the present invention on different datasets;

[0025] Figure 4 A schematic diagram of a face recognition device based on deep semantic perception provided in an embodiment of the present invention;

[0026] Figure 5 This is a partial structural block diagram of a face recognition terminal based on deep semantic perception provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, this embodiment of the invention employs a face recognition method based on deep semantic perception, including:

[0029] S1. Obtain the face image training set and construct the visual feature matrix of the face image training set. ;

[0030] Constructing the visual feature matrix of the face image training set This includes: flattening each face image into a one-dimensional vector to obtain the visual feature vector of each face image; and combining the visual feature vectors of all face images to obtain the visual feature matrix of the face image training set. .

[0031] S2, Visual feature matrix of the face image training set. Perform singular value decomposition to obtain the first-order standard projection matrix. ;

[0032] Visual feature matrix of the face image training set Perform singular value decomposition to obtain the first-order standard projection matrix. Includes: visual feature matrix of the face image training set Perform singular value decomposition to obtain the left singular value matrix, the diagonal matrix, and the right singular value matrix; set a threshold. Select the first left singular value matrix The singular vectors form a first-order standard projection matrix. Where D is the dimension of the visual feature vector of the face image.

[0033] threshold Based on the ablation experiment, the experiment showed that... The value is The best results were achieved during this period. This experiment was conducted on an AR dataset, using FRR (Face Recognition Accuracy) as the metric. The experimental data consisted of noisy face data, and the objective was to validate different... The value reflects the effectiveness of the method of this invention. The original dimension of the AR dataset was 19800, and it was found that... This method works best when the time is right, as analyzed below: When When the size is small, the range of the first-order projection subspace is small, resulting in excessive constraints on the second-order projection process. This means the method lacks sufficient degrees of freedom to acquire enough subspace information during second-order projection; when... When the size is large, the first-order projection subspace does not provide sufficient constraints for the second-order projection, causing the second-order projection process to deviate from the data manifold itself (i.e., deviate from the face space). Therefore, a suitable... It is essential to improve model efficiency.

[0034] The goal of this step is to obtain a geometrically interpretable subspace skeleton in the input space, providing a stable starting point for subsequent adaptive projections while preserving reconfigurability.

[0035] S3. Introducing a learnable matrix Constructing a second-order projection matrix Construct a two-stage projection reconstruction matrix Obtain a pre-trained face representation network Based on pre-trained face representation network Based on the visual feature matrix and two-stage projection reconstruction matrix Constructing the loss function ;

[0036] like Figure 2 As shown, the pre-trained face representation network The reconstruction residual is measured in the deep semantic space. This step transfers the "reconstruction-discrimination" consistency from the shallow pixel domain to the high-level semantic domain, avoiding the "optimal reconstruction but semantic mismatch" that occurs in traditional methods. To ensure fair comparison and engineering reusability, the pre-trained face representation network uniformly adopts mature backbone networks such as FaceNet in both experiments and implementation, serving only as a feature extractor and not requiring fine-tuning, ensuring that different algorithms can be compared and deployed on the same semantic foundation.

[0037] To explain the effectiveness of the method of this invention from a theoretical analysis perspective, at each sample... By applying a first-order Taylor approximation, we can obtain a quadratic approximation of the semantic residuals. Therefore, it can be seen that the residual energy is affected by the local sensitivity matrix. Weighting: Directional contributions insensitive to perceptual mapping are suppressed, while semantic directions relevant to discrimination are emphasized; from a statistical perspective, this process is equivalent to applying a weighted average to the perceptual manifold. and Error assessment is performed using co-induced Mahalanobis metrics, which naturally suppresses characteristic-level noise such as occlusion, illumination fluctuations, and background disturbances.

[0038] S4. Based on the loss function Iterative solution to find the optimal learnable matrix Based on the optimal learnable matrix Obtain the optimal second-order projection matrix ;

[0039] Two-stage projection consists of first-order projection and second-order projection. First-order projection projects the original data into a low-dimensional subspace, while second-order projection projects the original data into an even lower-dimensional subspace (which is inside the subspace obtained by first-order projection). In other words, first-order projection naturally constrains second-order projection, preventing the lower-dimensional subspace obtained by direct second-order projection from deviating from the range of the first-order projection subspace (this range is often strongly correlated with the original data, such as faces). Therefore, this constrained optimization effectively avoids overfitting on noisy samples and maintains stable generalization across datasets.

[0040] According to the loss function Iterative solution to find the optimal learnable matrix Including: in optimizing only And fixed Within this framework, the chain rule is used to calculate sequentially. , And synthesize the gradient in matrix form. Then, based on the learning rate renew And based on the threshold Determining convergence; specific steps include:

[0041] S41. Constructing the loss function Regarding learnable matrices gradient Initialize the learnable matrix ;

[0042] Constructing the loss function Regarding learnable matrices gradient include:

[0043] S411. Calculate the loss function about gradient :

[0044]

[0045] Where tr represents the trace of the matrix, i and j represent the row and column numbers of the matrix, respectively, and T represents the transpose;

[0046] S412. Calculate the loss function about gradient :

[0047]

[0048] S413, Calculation about gradient :

[0049]

[0050] in, These are the standard basis vectors, i.e. The i-th position is 1, and all other positions are 0;

[0051] S414, the loss function about gradient and about gradient Substitute into the loss function about gradient The final loss function is obtained. about gradient :

[0052]

[0053]

[0054] in, , .

[0055] S42, Freezing the pre-trained face representation network Using gradient descent based on the gradient For the current learnable matrix The updated learnable matrix is ​​obtained by performing an update. ;

[0056] Using gradient descent based on the gradient For the current learnable matrix The updates include: ;in, This is the learning rate.

[0057] S43. Based on the learnable matrix before and after the update , Calculate the loss function value respectively Determine whether it satisfies If this condition is met, then the optimal learnable matrix is ​​obtained. Otherwise, return to step S42; where k represents the number of iterations. This is the convergence threshold.

[0058] S5. Obtain the face image to be identified, and apply the optimal second-order projection matrix. Visual feature matrix of the face image training set The face image to be identified is then processed to obtain the face recognition result;

[0059] The process of recognizing a face image includes:

[0060] Constructing the visual feature vector of the face image to be identified Based on the visual feature vector of the face image to be identified and the optimal second-order projection matrix Calculate the low-dimensional representation of the face image to be identified. ;

[0061] Based on the visual feature matrix and the optimal second-order projection matrix Calculate the low-dimensional representation of each face image in the face image training set. ;in, It is the visual feature vector of the face images in the face image training set;

[0062] Based on the low-dimensional representation of the face image to be identified The nearest neighbor metric is used to make recognition decisions based on the low-dimensional representation of each face image in the face image training set, and the face recognition result of the face image to be recognized is obtained.

[0063] The recognition decision using the nearest neighbor metric includes: calculating the distance between the low-dimensional representation of the face image to be recognized and the low-dimensional representation of each face image in the face image training set; selecting the closest face image in the training set; and directly determining the category of the face image to be recognized based on the category of the closest image.

[0064] This reasoning approach, along with the semantic consistency target closed loop during training, enables reconstruction and discrimination to share the same metric basis, thereby maintaining stable recognition performance and interpretability in noisy scenarios.

[0065] The method of this invention was used to reconstruct different datasets, and the experimental results are as follows: Figure 3As shown in the figure, the present invention (DP-RPCA) compares with eight other similar methods (traditional PCA, L1-PCA, R1-PCA, RPCA-OM, Angle-PCA, T-RPCA, RPCA-AN, and N-RPCA) on six different standard datasets (AR, EYB, LFW, FERET, UMIST, and Multi-PIE). The horizontal axis represents the reduced dimension (compression level; a smaller value indicates greater compression), and the vertical axis represents the dimension-wise reconstruction loss. A lower curve indicates a smaller reconstruction error, meaning the compressed and restored image is more similar to the original. Simultaneously, the present invention was also tested in face recognition, and the experimental results are shown in Table 1, with the best results indicated in bold. It can be seen that the method used in the present invention improves the recognition rate compared to other common methods.

[0066] Table 1. Face recognition accuracy of different methods on different datasets (average of ten trials).

[0067]

[0068] The above experiments and related results analysis verify the effectiveness of the explanation method proposed in this invention.

[0069] like Figure 4 As shown, this embodiment of the invention also employs a face recognition device based on deep semantic perception to implement the above-mentioned face recognition method based on deep semantic perception, including:

[0070] The visual feature matrix construction module is used to obtain a training set of face images and construct the visual feature matrix of the training set of face images. ;

[0071] The orthogonal basis construction module is used to construct the visual feature matrix of the face image training set. Perform singular value decomposition to obtain the first-order standard projection matrix. That is, orthogonal basis;

[0072] The orthogonal basis construction module registers the obtained orthogonal basis as a read-only parameter for use in the second stage of projection learning.

[0073] The second projection and reconstruction module is used to introduce learnable matrices into a restricted subspace of an orthogonal basis. Constructing a second-order projection matrix Construct a two-stage projection reconstruction matrix ;

[0074] The perceptual consistency optimization module is used to obtain the pre-trained face representation network. Based on pre-trained face representation network Based on the visual feature matrix and two-stage projection reconstruction matrix Constructing a loss function within the deep semantic space According to the loss function Iterative solution to find the optimal learnable matrix Based on the optimal learnable matrix Obtain the optimal second-order projection matrix ;

[0075] The perceptual consistency optimization module can optionally weight the residuals based on feature sensitivity to emphasize directions relevant to identity determination and suppress perturbation directions that are insensitive to perceptual mapping, thereby improving robustness in occlusion, illumination changes and noisy scenarios.

[0076] The recognition output module is used to identify the optimal second-order projection matrix. Visual feature matrix of the face image training set The face image to be identified is then processed to obtain the face recognition result.

[0077] This module constructs the visual feature vector of the face image to be identified. Based on the visual feature vector of the face image to be identified and the optimal second-order projection matrix Calculate the low-dimensional representation of the face image to be identified; based on the visual feature matrix... and the optimal second-order projection matrix Calculate the low-dimensional representation of all face images in the face image training set; perform face recognition using the nearest neighbor metric based on the low-dimensional representations of all face images in the face image training set and the low-dimensional representation of the face image to be recognized, and obtain the face recognition result of the face image to be recognized.

[0078] This invention also provides a terminal, such as... Figure 5 As shown, for ease of explanation, only the parts related to the embodiments of the present invention are illustrated. The terminal may include any terminal device such as a mobile phone, tablet, computer, or POS machine; a computer is used as an example. Figure 5 The diagram shown is a block diagram of a portion of the structure of a computer related to the terminal provided in an embodiment of the present invention. The computer includes components such as a memory, input devices, output devices, an audio module, a WiFi module, a processor, and a power supply. Those skilled in the art will understand that... Figure 5 The computer architecture shown does not constitute a limitation on the computer and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0079] The memory is used to store the computer program for the face recognition method based on deep semantic perception described above;

[0080] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on computer usage, etc. In this invention, the data storage area may store pre-trained network parameters, orthogonal bases, and learnable matrix parameters. The memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other solid-state storage device.

[0081] A processor is used to run computer programs stored in memory, thereby performing various functions and data processing of the computer.

[0082] The processor may include one or more processing units, or it may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may also not be integrated into the processor.

[0083] An input device is used to receive input numeric or character information and to generate key signal inputs related to user settings and function control of a computer. The input device in this invention can be used to receive face image data or parameter configurations to be processed.

[0084] An output device is used to output and display information input by the user or information provided to the user, as well as various menus of the computer. The output device in this invention can be used to output the recognition results of a face image.

[0085] Output devices may include a display panel, which may be configured with a liquid crystal display or an organic light-emitting diode (OLED).

[0086] The audio module is used to convert the received audio data into electrical signals and transmit them to the speaker, where the speaker converts them into sound signals for output.

[0087] A WiFi module is a short-range wireless transmission technology. Computers can use WiFi modules to help users send and receive emails, browse web pages, and access streaming media, providing users with wireless broadband internet access.

[0088] The power source (such as a battery) is connected to the processor logic through a power management system, which then manages functions such as charging, discharging, and power consumption.

[0089] Although not in Figure 5 As shown in the diagram, the computer may also include a camera, Bluetooth module, etc., which will not be elaborated here.

[0090] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A face recognition method based on deep semantic perception, characterized in that, Comprising: S1, acquire a face image training set, construct a visual feature matrix of the face image training set ; S2, visual feature matrix of the face image training set singular value decomposition is performed to obtain a first-order standard projection matrix ; S3, introducing a learnable matrix constructing a second-order projection matrix constructing a two-stage projection reconstruction matrix obtaining a pre-trained face representation network based on the pre-trained face representation network according to the visual feature matrix and the two-stage projection reconstruction matrix constructing a loss function ; S4, according to the loss function iteratively solving for an optimal learnable matrix , according to the optimal learnable matrix obtaining an optimal second order projection matrix ; S5、acquire the face image to be identified, and identify the face image to be identified according to the optimal second-order projection matrix and the visual feature matrix of the face image training set identify the face image to be identified, and obtain a face recognition result. 2.The method of claim 1, wherein, Constructing a visual feature matrix of a face image training set The method comprises the following steps: flattening each face image into a one-dimensional vector respectively to obtain a visual feature vector of each face image; and combining the visual feature vectors of all face images to obtain a visual feature matrix of the face image training set . 3.The method of claim 1, wherein, The visual feature matrix of the training set of face images singular value decomposition to obtain a first-order standard projection matrix The visual feature matrix of the training set of face images singular value decomposition to obtain a left singular value matrix, a diagonal matrix and a right singular value matrix; set a threshold , select the first singular vectors of the left singular value matrix to form a first-order standard projection matrix .

4. The method of claim 1, wherein the method is based on deep semantic perception. According to the loss function Iteratively solving for the optimal learnable matrix Comprising: S41, constructing a loss function on the learnable matrix gradient ; initializing the learnable matrix ; S42, freeze the pre-trained face representation network , update the current learnable matrix according to the gradient by using the gradient descent method, to obtain an updated learnable matrix ; S43、According to the updated and the updated learnable matrix 、 Respectively calculate the loss function value , determine whether to meet , if so, get the optimal learnable matrix ; Otherwise, go back to step S42; wherein, k represents the number of iterations, is the convergence threshold.

5. The method of claim 4, wherein the method is based on deep semantic perception. Constructing a loss function Regarding the learnable matrix Gradient Comprise: S411、calculating a loss function with respect to the gradient of ; wherein tr denotes the trace of a matrix, i, j denote the row number and column number of a matrix, respectively, and T denotes the transpose; S412、compute the loss function about gradient : S413、calculating about gradient : S414, the loss function with respect to the gradient of and with respect to the gradient of substitute into the loss function with respect to the gradient of , resulting in the final loss function with respect to the gradient of .

6. The method of claim 5, wherein the method is based on deep semantic perception. computing a loss function with respect to the gradient of comprising: 。 7. The method of claim 6, wherein the method is based on deep semantic perception. Computing With respect to the gradient of comprising: wherein is the standard basis vector. 8.The method of claim 7, wherein, Final loss function Regarding Gradient of : wherein , . 9.The method of claim 1, wherein, According to the optimal second-order projection matrix and the face image training set identifies the face image to be identified, comprising: constructing a visual feature vector of a face image to be recognized , computing a low-dimensional representation of the face image to be recognized from the visual feature vector of the face image to be recognized and the optimal second-order projection matrix ; According to the visual feature matrix and the optimal second-order projection matrix Calculate the low-dimensional representation of each face image in the face image training set; According to the low-dimensional representation of the face image to be identified and the low-dimensional representation of each face image in the face image training set, a recognition decision is made by using a nearest neighbor metric to obtain a face recognition result of the face image to be identified.

10. A deep semantic perception based face recognition system characterized in that, A face recognition method according to any one of claims 1-9, comprising: A face recognition method according to any one of claims 1-9, comprising: The visual feature matrix construction module is configured to acquire the face image training set and construct a visual feature matrix of the face image training set ; An orthogonal basis construction module is configured to perform singular value decomposition on the visual feature matrix of the training set of face images to obtain a first-order standard projection matrix ; a secondary projection and reconstruction module for introducing a learnable matrix constructing a second order projection matrix , constructing a two-stage projection reconstruction matrix ; The perception consistency optimization module is configured to obtain a pre-trained face representation network based on the pre-trained face representation network according to a visual feature matrix and a two-stage projection reconstruction matrix to construct a loss function according to the loss function iteratively solve an optimal learnable matrix according to the optimal learnable matrix obtain an optimal second-order projection matrix ​ The recognition output module is configured to recognize the face image to be recognized according to the optimal second-order projection matrix and the visual feature matrix of the face image training set and obtain a face recognition result.

Images