Low-resolution face recognition method and system based on frequency domain feature knowledge distillation

Abstract

The invention discloses a low-resolution face recognition method and system based on frequency domain feature knowledge distillation. The method comprises the steps that a teacher model and a student model are constructed, the teacher model and the student model both adopt the same FRENet frequency domain network structure, the structure is composed of N cascaded FREblock frequency domain modules, and each FREblock comprises a spatial domain branch and a frequency domain branch; training a teacher model by using a high-resolution data set, and calculating classification loss as supervision; training a student model by using a low-resolution data set, calculating classification loss as supervision, calculating cosine similarity between frequency domain features output by FREblocks of teachers and students as frequency domain feature loss, and calculating KL divergence between output probabilities of the teachers and the students as knowledge distillation loss; and jointly optimizing the three losses to obtain a final student model for low-resolution face recognition. The student model can fully learn the frequency domain feature distribution rule of the teacher model, and the face recognition accuracy is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a low-resolution face recognition method and system based on frequency domain feature knowledge distillation. BACKGROUND

[0002] Face recognition, as one of the core technologies in the field of computer vision, has made significant progress in recent years, especially under the promotion of deep learning technology. At present, high-resolution (HR) face recognition technology has made breakthrough progress in multiple sub-directions, such as cross-modal recognition, cross-age recognition, and face recognition of specific groups, and is widely used in security, finance, social media and other practical scenarios. However, in real-world applications, the collected face images often have complex environmental interference, such as long-distance shooting, insufficient lighting, occlusion or motion blur, etc. These factors usually cause the image resolution to decrease significantly and make the key discriminative features degrade.

[0003] Since most of the existing mainstream face recognition methods rely on high-quality images for training, their performance often decreases significantly when facing low-resolution (LR) images. Therefore, how to improve the ability of the model to extract face features under the LR condition has become one of the key challenges to improve the practicality of face recognition. Researchers have proposed a variety of LR face recognition techniques, which can be roughly divided into two categories: reconstruction-based methods and embedding-based methods.

[0004] Reconstruction-based methods primarily focus on recovering a version of the original high-resolution image from a low-resolution (LR) image, typically employing super-resolution (SR) reconstruction techniques to enhance visual quality. These methods utilize strategies such as Generative Adversarial Networks (GANs), residual blocks, or perceptual losses to make the output image structurally and texturally clearer. While these methods visually improve image quality, they don't necessarily recover key discriminative features helpful for identity recognition, thus their performance improvement is often limited, especially in cases of blurred or severely missing information. In contrast, embedding-based methods directly learn discriminative features from LR images in the feature space. Some embedding methods attempt to improve the discriminative power of low-resolution features by directly designing structures and loss functions without relying on any teacher model. These methods typically employ specific network structures to enhance the perception of low-resolution images, such as introducing attention mechanisms, adaptive convolutional modules, or enhancing feature representation through multi-scale feature fusion and graph structure modeling. However, due to the lack of detailed information in LR images, directly learning sufficiently discriminative features is quite challenging.

[0005] To compensate for the lack of image information, a more effective solution is to introduce auxiliary information from the HR image to guide the feature learning process. The goal of this type of method is to make the projection of the LR image in the embedding space as close as possible to the projection of the corresponding HR image, thereby achieving consistent recognition across resolutions. This idea has led to the widespread application of Knowledge Distillation (KD) technology in LR face recognition. The core idea of ​​KD is to use a teacher model trained on HR images as a knowledge source, transferring its knowledge in classification, feature space projection, and sample relationship modeling to a student model that takes LR images as input. In this way, the student model can acquire discriminative knowledge from the teacher model during training, thereby effectively improving its ability to recognize low-resolution images.

[0006] Common KD methods include soft label distillation, feature alignment, and relational knowledge transfer. These methods have all been proven in practice to improve accuracy to varying degrees. However, most existing KD methods are based on Convolutional Neural Networks (CNNs). As network depth increases, the computational resources required to expand the receptive field grow exponentially, making the model inefficient when modeling long-range dependencies. Furthermore, some studies have attempted to introduce Transformer-based structures for distillation, using self-attention mechanisms for global feature modeling. However, because Transformers lack inherent local inductive biases, they struggle to effectively extract local details and boundary features in low-resolution scenes, leading to unstable recognition performance.

[0007] To effectively address the aforementioned problems, this invention proposes a low-resolution face recognition method based on frequency domain feature knowledge distillation. The core of this method lies in the modeling and efficient transfer of global frequency domain features. Summary of the Invention

[0008] To address the above technical problems, this invention provides a low-resolution face recognition method and system based on frequency domain feature knowledge distillation.

[0009] The technical solution adopted by this invention to solve its technical problem is: A low-resolution face recognition method based on frequency domain feature knowledge distillation includes the following steps: S100: Downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset; S200: Construct teacher and student models, both of which use the same FRENet frequency domain network structure. FRENet consists of N cascaded FREblock frequency domain modules, and each FREblock contains spatial domain branches and frequency domain branches. S300: First, shallow features of the original input image are extracted through single-layer convolution. The spatial domain branch extracts the spatial features of the shallow features. The frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. After fusing the spatial features and frequency domain features, they are connected with the residual of the original input features as the original input features of the next FREblock. S400: Train the teacher model using a high-resolution face dataset and compute ArcFace loss as classification supervision; S500: Train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model in each FREblock output as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. S600: Jointly optimize the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to obtain the final student model, and complete low-resolution face recognition based on the final student model.

[0010] Preferably, the S300 mid-frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform, specifically as follows: The shallow features are decomposed into different frequency subbands, HH, HL, LH, and LL, using Discrete Wavelet Transform (DWT), as follows: ; in, For high-frequency sub-band, For low-frequency sub-band, This is a shallow feature.

[0011] Preferably, the processing of the high-frequency subband in S300 includes: In the high-frequency subband, the three high-frequency components HH, HL, and LH are first spliced ​​together. After normalization, the features are then re-fused and weighted using a BasicBlock residual module, and the number of channels is adjusted to obtain the high-frequency features. Specifically: ; in, This indicates a channel splicing operation. This indicates a normalization operation.

[0012] Preferably, the processing of the low-frequency subband in S300 includes: For the low-frequency subband LL, it is first transformed to the frequency domain using FFT, for a given input image. The FFT operation is mathematically represented as: ; in, Represents the c-th channel and frequency in the frequency domain. The result of the two-dimensional discrete Fourier transform at the given location; c represents the channel index. and These represent the height and width of the spatial domain image, respectively; Represents the c-th channel and its position in the spatial domain. Pixel value at that location, and These are the pixel coordinates in the spatial domain, specifically the pixel indices in the vertical and horizontal directions. and These are frequency component indices in the frequency domain, specifically the vertical frequency index and the horizontal frequency index. It is a complex exponential function; For the obtained frequency domain features, a learnable filter is applied to weight different components, adjusting the importance of different information; assuming the number of input feature channels is... The number of output channels is The global filter adopts The number of filter kernels processes the input features, which can be mathematically represented as: ; in, Represents the input frequency domain feature tensor. This represents a learnable filter bank. Represents the output frequency domain feature tensor; The filtered features are then transformed back into the spatial domain using IFFT. The IFFT operation is mathematically represented as: ; ; in, This represents the frequency domain characteristics after filtering. This represents the spatial domain characteristics of the output after IFFT transformation. is the complex exponential basis function of the inverse Fourier transform. This is for channel splicing operations. The spatial domain feature maps of each channel after IFFT transformation. This represents the final low-frequency characteristics obtained after splicing.

[0013] Preferably, in S300, the high-frequency and low-frequency processing results are fused into frequency domain features, specifically as follows: The high-frequency output branch and the low-frequency output branch are fused by the Fusionblock fusion module. The Fusionblock consists of two Convblock convolutional blocks. The final fusion process is represented as follows: ; in, High-frequency characteristics It is a low-frequency characteristic. For frequency domain characteristics, For spatial features, For the fusion output characteristics of FREblock, This represents the activation function. This indicates a convolutional layer.

[0014] Preferably, in S300, after fusing spatial features and frequency domain features, the result is concatenated with the residual of the original input features and used as the original input features for the next FREblock, specifically: ; in, These are the original input features for the next FRE block. These are the original input features.

[0015] Preferably, in S400, the teacher model training uses ArcFace as the classification loss, specifically: ; in, Scaling factor For edge threshold, This represents the angle between the sample and the true class center. This represents the angle between the sample and the j-th class.

[0016] Preferably, the frequency domain feature loss in S500 is calculated as follows: Suppose the frequency domain characteristics of the teacher model in the nth FRE block stage are represented as follows: The output of the student model at this stage is represented as For the nth FRE block, the frequency domain feature loss The cosine similarity is expressed as the frequency domain feature vectors of the teacher model and the student model, specifically: ; Overall frequency domain feature loss Represented as: ; Where N is the total number of FREblocks; Knowledge distillation loss in S500 The KL divergence softened by the temperature coefficient T is as follows: ; Where T represents the temperature coefficient of knowledge distillation loss, used for the softening probability distribution. The softened probability distribution for the teacher model. The softened probability distribution for the student model. Let be the softening probability value of the teacher model in the i-th category. Let be the softening probability value of the student model in the i-th category.

[0017] Preferably, the S600 jointly optimizes the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss as follows: ; in , representing the overall weight of the classification loss.

[0018] A low-resolution face recognition system based on frequency domain feature knowledge distillation includes: The downsampling module is used to downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset. The model building module is used to build teacher and student models. Both of them use the same FRENet frequency domain network structure. FRENet consists of N cascaded FREblock frequency domain modules. Each FREblock contains spatial domain branches and frequency domain branches. The spatial and frequency domain feature extraction modules extract spatial features from the original input features through the spatial domain branch, and decompose the original input features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform through the frequency domain branch. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. The spatial features and frequency domain features are then fused and connected with the residual of the original input features as the original input features for the next FREblock. The teacher model training module is used to train a teacher model using a high-resolution face dataset and calculates the ArcFace loss as classification supervision. The student model training and distillation module is used to train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model output in each FREblock as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. Finally, a joint optimization module is used to jointly optimize the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to obtain the final student model. Low-resolution face recognition is then performed based on the final student model.

[0019] The aforementioned low-resolution face recognition method and system based on frequency domain feature knowledge distillation innovatively integrates a global frequency domain feature extraction component into the ResNet architecture, proposes FREblock, and constructs a novel network structure, FRENet. This network outputs discriminative frequency domain features at each processing stage. Through the knowledge distillation framework, the student model can fully learn the high-quality frequency domain feature representations provided by the teacher model, thereby significantly improving its feature extraction capability under low-resolution conditions. Attached Figure Description

[0020] Figure 1 This is a flowchart of a low-resolution face recognition method based on frequency domain feature knowledge distillation in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall model structure of a low-resolution face recognition method based on frequency domain feature knowledge distillation in one embodiment of the present invention. Figure 3 This is a schematic diagram of FREblock in one embodiment of the present invention; Figure 4 This is a schematic diagram of FusionBlock in one embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the clustering effect of LR faces in one embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] In one embodiment, such as Figure 1 and 2 As shown, a low-resolution face recognition method based on frequency domain feature knowledge distillation includes the following steps: S100: Downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset; S200: Construct teacher and student models, both of which use the same FRENet (Frequency Enhanced Network) frequency domain network structure. FRENet consists of N cascaded FREblock (Frequency Enhancement Block) frequency domain modules, and each FREblock contains spatial domain branches and frequency domain branches. S300: First, shallow features of the original input image are extracted through single-layer convolution. The spatial domain branch extracts the spatial features of the shallow features. The frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. After fusing the spatial features and frequency domain features, they are connected with the residual of the original input features as the original input features of the next FREblock. S400: Train the teacher model using a high-resolution face dataset and compute ArcFace loss as classification supervision; S500: Train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model in each FREblock output as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. S600: Jointly optimize the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to obtain the final student model, and complete low-resolution face recognition based on the final student model.

[0023] Specifically, the purpose of this invention is to address the limitations of existing low-resolution face recognition methods in frequency domain feature modeling. Traditional methods mainly rely on spatial domain feature learning, which struggles to effectively capture the global frequency domain distribution characteristics of face images, leading to a sharp decline in recognition performance when image resolution is significantly reduced. To address this, this invention proposes a low-resolution face recognition method based on frequency domain feature knowledge distillation. It innovatively integrates a global frequency domain feature extraction module into the ResNet (Residual Network) architecture, introduces FREblock, and constructs a novel network structure, FRENet. Through multi-stage frequency domain feature distillation, the student model can fully learn the multi-level frequency domain feature representations provided by the teacher model, thereby significantly improving the model's feature discrimination ability and recognition robustness under low-resolution conditions.

[0024] In one embodiment, such as Figure 3 As shown, the spatial domain branch of FREblock is composed of the residual module Basicblock, which extracts spatial features.

[0025] Specifically, the spatial branch follows the stacked BasicBlock module approach in traditional ResNet. In order to achieve better complementarity with the frequency domain features, through experiments, this invention stacks 4 BasicBlock modules in the spatial branch for feature extraction.

[0026] In one embodiment, such as Figure 3 As shown, the S300 mid-frequency domain branch decomposes the shallow features into high-frequency and low-frequency sub-bands through discrete wavelet transform, specifically: The shallow features are decomposed into different frequency subbands HH, HL, LH, and LL using Discrete Wavelet Transform (DWT), specifically as follows: ; in, For high-frequency sub-band, For low-frequency sub-band, This is a shallow feature.

[0027] In one embodiment, processing of the high-frequency subband in S300 includes: In the high-frequency subband, the three high-frequency components HH, HL, and LH are first spliced ​​together. After normalization, the features are then re-fused and weighted using a BasicBlock residual module, and the number of channels is adjusted to obtain the high-frequency features. Specifically: ; in, This indicates a channel splicing operation. This indicates a normalization operation.

[0028] In one embodiment, processing the low-frequency subband in S300 includes: For the low-frequency subband LL, it is first transformed to the frequency domain using FFT, for a given input image. The FFT operation is mathematically represented as: ; in, Represents the c-th channel and frequency in the frequency domain. The result of the two-dimensional discrete Fourier transform at the given location; c represents the channel index. and These represent the height and width of the spatial domain image, respectively; Represents the c-th channel and its position in the spatial domain. Pixel value at that location, and These are the pixel coordinates in the spatial domain, specifically the pixel indices in the vertical and horizontal directions. and These are frequency component indices in the frequency domain, specifically the vertical frequency index and the horizontal frequency index. It is a complex exponential function; For the obtained frequency domain features, a learnable filter is applied to weight different components, adjusting the importance of different information; assuming the number of input feature channels is... The number of output channels is The global filter adopts The number of filter kernels processes the input features, which can be mathematically represented as: ; in, Represents the input frequency domain feature tensor. This represents a learnable filter bank. Represents the output frequency domain feature tensor; The filtered features are then transformed back into the spatial domain using IFFT. The IFFT operation is mathematically represented as: ; ; in, This represents the frequency domain characteristics after filtering. This represents the spatial domain characteristics of the output after IFFT transformation. is the complex exponential basis function of the inverse Fourier transform. This is for channel splicing operations. The spatial domain feature maps of each channel after IFFT transformation. This represents the final low-frequency characteristics obtained after splicing.

[0029] In one embodiment, S300 fuses the high-frequency and low-frequency processing results into frequency domain features, specifically as follows: The high-frequency output branch and the low-frequency output branch are fused by the Fusionblock fusion module. The Fusionblock consists of two Convblock convolutional blocks. The final fusion process is represented as follows: ; in, High-frequency characteristics It is a low-frequency characteristic. For frequency domain characteristics, For spatial features, For the fusion output characteristics of FREblock, This represents the activation function. This indicates a convolutional layer.

[0030] In one embodiment, after fusing spatial features and frequency domain features in S300, the resulting feature is concatenated with the residual of the original input features and used as the original input features for the next FRE block. Specifically: ; in, These are the original input features for the next FRE block. These are the original input features.

[0031] In one embodiment, the teacher model training in S400 uses ArcFace as the classification loss, specifically: ; in, Scaling factor For edge threshold, This represents the angle between the sample and the true class center. This represents the angle between the sample and the j-th class.

[0032] Specifically, in this embodiment, Set to 32, edge distance Set it to 0.5.

[0033] In one embodiment, the frequency domain feature loss is calculated by plotting the angle between the frequency domain features of the teacher model and the student model in angular space. Ideally, the angular space distributions of the two models should be as similar as possible; this is achieved through multi-stage loss calculation. For example... Figure 2As shown, the frequency domain feature loss of S500 is calculated as follows: Suppose the frequency domain characteristics of the teacher model in the nth FRE block stage are represented as follows: The output of the student model at this stage is represented as For the nth FRE block, the frequency domain feature loss The cosine similarity is expressed as the frequency domain feature vectors of the teacher model and the student model, specifically: ; Overall frequency domain feature loss Represented as: ; Where N is the total number of FREblocks; Knowledge distillation loss helps the student model learn the probability distribution of the teacher model, typically achieved using KL divergence. In S500, knowledge distillation loss... The KL divergence softened by the temperature coefficient T is as follows: ; Where T represents the temperature coefficient of knowledge distillation loss, used for the softening probability distribution. The softened probability distribution for the teacher model. The softened probability distribution for the student model. Let be the softening probability value of the teacher model in the i-th category. Let be the softening probability value of the student model in the i-th category.

[0034] Specifically, in this embodiment, FRENet is obtained by connecting 4 FREblocks.

[0035] In one embodiment, S600 jointly optimizes the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to find the most suitable weight distribution, specifically as follows: ; in , representing the overall weight of the classification loss. In this embodiment, Set it to 0.7.

[0036] In one instance, such as Figure 5 As shown, a total of 10 identities were selected, each containing 50 different face images. The results of clustering these 500 low-resolution face images by category (using the t-SNE method) are obtained. As can be seen from the clustering results, the method proposed in this invention can cluster different identities well, and there is a clear boundary distance between each category, indicating that the method of this invention has superiority in low-resolution face feature extraction.

[0037] The aforementioned low-resolution face recognition method based on frequency domain feature knowledge distillation has the following beneficial effects: (1) The low-resolution face recognition method based on frequency domain feature knowledge distillation proposed in this invention can effectively improve the student model's ability to perceive global structural information and local details, thereby maintaining a high face recognition accuracy even when the image quality is severely degraded. This method has the advantages of lightweight structure and efficient training, and is suitable for practical deployment scenarios, especially in edge computing and mobile devices, where it has good application potential.

[0038] (2) The designed FREblock, by introducing two frequency domain representations, Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT), realizes low-resolution global feature modeling and high-frequency local texture feature extraction of face images. It can extract richer and more complementary feature representations at different frequencies, effectively enhancing the discriminative ability of feature expression. The proposed FRENet is composed of FREblocks through residual connections.

[0039] (3) The proposed frequency domain feature distillation loss function achieves efficient transfer of structural knowledge by aligning the angles of the teacher model and the student model in the frequency domain feature space. This enables the student model to fully learn the frequency domain feature distribution law of the teacher model, improves the stability and effectiveness of the distillation process, and helps to improve the overall recognition performance.

[0040] In one embodiment, a low-resolution face recognition system based on frequency domain feature knowledge distillation is also provided, comprising: The downsampling module is used to downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset. The model building module is used to build teacher and student models. Both of them use the same FRENet frequency domain network structure. FRENet consists of N cascaded FREblock frequency domain modules. Each FREblock contains spatial domain branches and frequency domain branches. The spatial and frequency domain feature extraction modules first extract shallow features from the original input image through single-layer convolution. The spatial domain branch extracts the spatial features of the shallow features, and the frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. The spatial features and frequency domain features are then fused and connected with the residual of the original input features as the original input features for the next FREblock. The teacher model training module is used to train a teacher model using a high-resolution face dataset and calculates the ArcFace loss as classification supervision. The student model training and distillation module is used to train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model output in each FREblock as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. Finally, a joint optimization module is used to jointly optimize the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to obtain the final student model. Low-resolution face recognition is then performed based on the final student model.

[0041] Specifically, in the teacher model training phase, the model is configured to use high-resolution face images as input and is trained using the designed FRENet. Identity supervision is performed using ArcFace loss, enabling the model to learn discriminative features with global structure and detail awareness. In the student model training and distillation phase, the model is configured to use low-resolution face images as input and is trained using a structurally consistent FRENet student model. While maintaining the ArcFace discriminative loss, a distillation loss based on frequency domain feature alignment is introduced to guide the student model to effectively mimic the teacher model's ability in frequency domain modeling. Finally, in the joint optimization phase, a weighted combination of discriminative ArcFace loss and frequency domain distillation loss is used as the joint optimization objective. Performance balance is achieved by adjusting the loss weights, thereby improving the student model's global modeling ability and feature generalization ability under low-quality input conditions.

[0042] Specific limitations regarding the low-resolution face recognition system based on frequency domain feature knowledge distillation can be found in the limitations of the low-resolution face recognition method based on frequency domain feature knowledge distillation described above, and will not be repeated here. Each module in the aforementioned low-resolution face recognition system based on frequency domain feature knowledge distillation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0043] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a low-resolution face recognition method based on frequency domain feature knowledge distillation.

[0044] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a low-resolution face recognition method based on frequency domain feature knowledge distillation.

[0045] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0046] The above provides a detailed description of a low-resolution face recognition method and system based on frequency domain feature knowledge distillation provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention, and the descriptions of the embodiments are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A low-resolution face recognition method based on frequency domain feature knowledge distillation, characterized in that, Includes the following steps: S100: Downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset; S200: Construct teacher and student models, both of which use the same FRENet frequency domain network structure. FRENet consists of N cascaded FREblock frequency domain modules, and each FREblock contains spatial domain branches and frequency domain branches. S300: First, shallow features of the original input image are extracted through single-layer convolution. The spatial domain branch extracts the spatial features of the shallow features. The frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. After fusing the spatial features and frequency domain features, they are connected with the residual of the original input features as the input features of the next FREblock. S400: Train the teacher model using a high-resolution face dataset and compute ArcFace classification loss as classification supervision; S500: Train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model in each FREblock output as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. S600: Jointly optimize frequency domain feature loss, ArcFace and knowledge distillation loss to obtain the final student model, and complete low-resolution face recognition based on the final student model.

2. The method according to claim 1, characterized in that, The S300 mid-frequency domain branch decomposes shallow features into high-frequency and low-frequency sub-bands using discrete wavelet transform, specifically: The shallow features are decomposed into different frequency subbands using the Discrete Wavelet Transform (DWT), including the HH high-frequency-high-frequency subband, the HL high-frequency-low-frequency subband, the LH low-frequency-high-frequency subband, and the LL low-frequency-low-frequency subband, specifically: ； in, For high-frequency sub-band, For low-frequency sub-band, This is a shallow feature.

3. The method according to claim 2, characterized in that, The S300 processes high-frequency subbands, including: In the high-frequency subband, the three high-frequency components HH, HL, and LH are first spliced ​​together. After normalization, the features are then re-fused and weighted using a BasicBlock residual module, and the number of channels is adjusted to obtain the high-frequency features. Specifically: ； in, This indicates a channel splicing operation. This indicates a normalization operation.

4. The method according to claim 3, characterized in that, The S300 processes low-frequency subbands, including: For the low-frequency subband LL, it is first transformed to the frequency domain using FFT (Fast Fourier Transform). This is then applied to an input image. The FFT operation is mathematically represented as: ； in, Represents the c-th channel and frequency in the frequency domain. The result of the two-dimensional discrete Fourier transform at the given location; c represents the channel index. and These represent the height and width of the spatial domain image, respectively; Represents the c-th channel and its position in the spatial domain. Pixel value at that location, and These are the pixel coordinates in the spatial domain, specifically the pixel indices in the vertical and horizontal directions. and These are frequency component indices in the frequency domain, specifically the vertical frequency index and the horizontal frequency index. It is a complex exponential function; For the obtained frequency domain features, a learnable filter is applied to weight different components, adjusting the importance of different information; assuming the number of input feature channels is... The number of output channels is The global filter adopts The number of filter kernels processes the input features, which can be mathematically represented as: ； in, Represents the input frequency domain feature tensor. This represents a learnable filter bank. Represents the output frequency domain feature tensor; The filtered features are then transformed back into the spatial domain using the inverse fast Fourier transform (IFFT). The IFFT operation is mathematically represented as: ； ； in, This represents the frequency domain characteristics after filtering. This represents the spatial domain characteristics of the output after IFFT transformation. is the complex exponential basis function of the inverse Fourier transform. This is for channel splicing operations. The spatial domain feature maps of each channel after IFFT transformation. This represents the final low-frequency characteristics obtained after splicing.

5. The method according to claim 4, characterized in that, In S300, the high-frequency and low-frequency processing results are fused into frequency domain features, specifically: The high-frequency output branch and the low-frequency output branch are fused by the Fusionblock fusion module. The Fusionblock consists of two Convblock convolutional blocks. The final fusion process is represented as follows: ； in, High-frequency characteristics It is a low-frequency characteristic. For frequency domain characteristics, For spatial features, For the fusion output features of FREblock, This represents the activation function. This indicates a convolutional layer.

6. The method according to claim 5, characterized in that, In S300, spatial features and frequency domain features are fused and then concatenated with the residual of the original input features to serve as the original input features for the next FRE block. Specifically: ； in, These are the original input features for the next FRE block. These are the original input features.

7. The method according to claim 6, characterized in that, In S400, the teacher model training uses ArcFace as the classification loss, specifically: ； in, Scaling factor For edge threshold, This represents the angle between the sample and the true class center. This represents the angle between the sample and the j-th class.

8. The method according to claim 7, characterized in that, The frequency domain feature loss of S500 is calculated as follows: Suppose the frequency domain characteristics of the teacher model in the nth FRE block stage are represented as follows: The output of the student model at this stage is represented as For the nth FRE block, the frequency domain feature loss The cosine similarity is expressed as the frequency domain feature vectors of the teacher model and the student model, specifically: ； Overall frequency domain feature loss Represented as: ； Where N is the total number of FREblocks; Knowledge distillation loss in S500 The KL divergence softened by the temperature coefficient T is as follows: ； Where T represents the temperature coefficient of knowledge distillation loss, used for the softening probability distribution. The softened probability distribution for the teacher model. The softened probability distribution for the student model. Let be the softening probability value of the teacher model in the i-th category. Let be the softening probability value of the student model in the i-th category.

9. The method according to claim 8, characterized in that, The S600 algorithm jointly optimizes the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss as follows: ； in , representing the overall weight of the classification loss.

10. A low-resolution face recognition system based on frequency domain feature knowledge distillation, characterized in that, include: The downsampling module is used to downsample the original high-resolution face image dataset to generate a corresponding low-resolution dataset. The model building module is used to build teacher and student models. Both of them use the same FRENet frequency domain network structure. FRENet consists of N cascaded FREblock frequency domain modules. Each FREblock contains spatial domain branches and frequency domain branches. The spatial and frequency domain feature extraction modules first extract shallow features from the original input image through single-layer convolution. The spatial domain branch extracts the spatial features of the shallow features, and the frequency domain branch decomposes the shallow features into high-frequency sub-bands and low-frequency sub-bands through discrete wavelet transform. After processing the high-frequency sub-bands and low-frequency sub-bands respectively, they are fused into frequency domain features. The spatial features and frequency domain features are then fused and connected with the residual of the original input features as the original input features for the next FREblock. The teacher model training module is used to train a teacher model using a high-resolution face dataset and calculates the ArcFace loss as classification supervision. The student model training and distillation module is used to train the student model using a low-resolution face dataset, calculate ArcFace loss as classification supervision, calculate the cosine similarity between the frequency domain features of the teacher model and the student model output in each FREblock as the frequency domain feature loss, and calculate the KL divergence between the output probabilities of the teacher model and the student model as the knowledge distillation loss. Finally, a joint optimization module is used to jointly optimize the frequency domain feature loss, ArcFace classification loss, and knowledge distillation loss to obtain the final student model. Low-resolution face recognition is then performed based on the final student model.

Images