CN117830096A - A super-resolution reconstruction method for face images under low light conditions - Google Patents

Abstract

The present invention discloses a method for super-resolution reconstruction of facial images under low-light conditions. Step 1: synthesize a low-light low-resolution facial image I _LLR ; Step 2: construct a brightness-corrected facial super-resolution network IC‑FSRNet; Step 3: input the image synthesized in step 1 into step 2, improve the brightness of the facial image and restore the facial structure information, and obtain I _SR1 ; Step 4: construct a detail enhancement model DENet; Step 5: input the image obtained in step 3 into step 4, improve the facial details of the facial image, so that the facial image has a better visual effect, and obtain I _SR2 . The present invention can effectively improve the visual quality of low-light low-resolution facial images and solve the problem of loss of important facial information in the existing cascade technology.

Description

A super-resolution reconstruction method for face images under low light conditions

Technical Field

The invention belongs to the field of images, and in particular relates to a super-resolution reconstruction method for face images under low-light conditions.

Background technique

Face super-resolution (FSR), also known as face hallucination, is a technique to recover a corresponding high-resolution face image from a low-resolution face image. Face super-resolution has been studied for many years and is widely used in outdoor computer vision systems. Although existing face super-resolution methods are very effective in processing face images captured under normal lighting conditions, there is still room for improvement under low-light conditions. In practical scenarios, due to the diversity and complexity of imaging environments (such as insufficient lighting at night or limited exposure time in video surveillance scenarios), the captured face images are simultaneously interfered by low resolution and low-light degradation. Therefore, there is an urgent need to design low-light low-resolution face image super-resolution techniques.

Before deep learning, traditional methods improve the resolution of face images by designing handcrafted priors and reconstruction operators, but their generalization ability is limited in open-world scenarios. Recently, deep convolutional neural networks have emerged and have shown superior performance on face super-resolution tasks than traditional methods. However, these methods still cannot reconstruct visually pleasing and content-preserving results from low-light low-resolution face images because they are tailored to restore face images captured under normal lighting conditions. One possible approach is to perform face super-resolution and low-light image enhancement (LLIE) sequentially, such as super-resolution first and then enhancement of the super-resolution result (FSR+LLIE) or image enhancement first and then super-resolution (LLIE+FSR). However, simple cascade solutions are difficult to generate results with satisfactory facial structure and visual quality. The reason is that individual methods only focus on a single degradation (e.g., low-light image enhancement methods only focus on low-light degradation, while face super-resolution methods only focus on low-resolution degradation), while the input face images suffer from both low-light and low-resolution degradation. That is, the former process will inevitably produce artifacts and errors, which will interfere with the latter process and cause error accumulation, ultimately leading to the loss of important facial information. Therefore, there is an urgent need for a method that can effectively restore low-light and low-resolution face images.

Summary of the invention

The present invention provides a super-resolution reconstruction method for facial images under low-light conditions, which is used to solve the problem in the prior art of error accumulation leading to loss of important facial information.

The present invention is achieved through the following technical solutions:

A method for super-resolution reconstruction of face images under low-light conditions, the reconstruction method comprising the following steps:

Step 1: Synthesize a low-light low-resolution face image I _LLR ;

Step 2: Construct the brightness correction face super-resolution network IC-FSRNet;

Step 3: Input the image synthesized in step 1 into step 2 to improve the brightness of the face image and restore the face structure information to obtain I _SR1 ;

Step 4: Build the detail enhancement model DENet;

Step 5: Input the image obtained in step 3 into step 4 to improve the facial details of the face image, so that the face image has a better visual effect, and obtain I _SR2 .

Furthermore, the said constructing brightness correction face super-resolution network specifically includes a brightness estimation IE branch for estimating a brightness coefficient for brightness adjustment and a face super-resolution FSR branch for reconstructing a super-resolution result;

First, the low-light low-resolution face image I _LLR is sent to two branches;

Then, the brightness correction super-resolution block ICSRB takes the output features of the previous step as input to encode their mutual relations.

Furthermore, the brightness correction super-resolution block ICSRB converts the output features of the previous step Specifically, three cascaded convolutional layers are used as input to encode their mutual relationship. Predict the bilateral brightness grid B ₁ ; at the same time, extract the guidance map G from the upsampled I _LLR through the convolution layer; with B ₁ and G, the brightness coefficient C ₁ is obtained through 3D slicing, expressed as,

C ₁ = f _Slice (B ₁ ,G),

Where f _Slice represents a 3D slice operation;

The brightness coefficient _C1 is then fed into the brightness adjustment block to adjust the brightness of the feature _S1 in the FSR branch, producing the adjusted result

Further, the 3D slice first projects the brightness grid B ₁ onto the 3D grid using G;

The mesh is then blurred using Gaussian blur;

Finally, according to the blurred bilateral grid and the guidance image G, the brightness coefficient C ₁ is obtained by accessing the grid values using trilinear interpolation.

Furthermore, in addition to using the brightness coefficient from the IE branch to improve the brightness in the FSR branch, the brightness adjustment result is also Feedback to the brightness refinement block of the IE branch to refine the original brightness and improve the subsequent brightness estimation; in this way, face super-resolution can promote brightness estimation; then, the generated super-resolution result and brightness refinement result are input into the following L-1 ICSRBs for more accurate mutual refinement and utilization;

After L times of brightness estimation and face super-resolution, a convolutional layer is applied to the result of the FSR branch to generate the reconstruction result I _SR1 ; in order to optimize the network, L1 pixel loss is used,

Among them, I _HR is the reference standard for high-quality face images.

Furthermore, the brightness adjustment is specifically to first map the learned coefficient _Ci to [-1,1] through the convolution layer and the Sigmoid function to obtain the brightness coefficient _C'i . The super-resolution feature _Si will be input into the cascaded convolution layer to generate _S'i , and then the brightness is adjusted through the following conversion function:

in represents the brightness-adjusted feature, and * indicates pixel-wise multiplication.

Furthermore, the brightness refinement is specifically as follows: given the super-resolution result and the original brightness _Bi , using different fully connected layers to generate query Q from _Bi , and from/> Get the key K and value V; then calculate the cross attention between them,

Where F _Att represents the attention map, f _Softmax represents softmax, and d is a hyperparameter;

Then, a feed-forward network is applied to generate the refined result; the super-resolution features are used to refine the brightness estimation in the next step.

Furthermore, the construction of the detail enhancement model DENet is specifically as follows:

Where ∈～N(0,1), (x,y) corresponds to I _SR1 and I _HR , γ～p(γ), and f _DENet represents DENet.

The beneficial effects of the present invention are:

The present invention divides the low-light face image super-resolution task into structure fidelity reconstruction and texture consistency learning, and designs a brightness correction face super-resolution network (IC-FSRNet) and a detail enhancement model (DENet); the former aims to reproduce face images with credible structural information, while the latter is responsible for eliminating disturbances to improve texture consistency.

The present invention makes full use of the complementary information between brightness correction and face super-resolution so that the two complement each other.

Our experimental results demonstrate that the proposed method achieves state-of-the-art performance in terms of both visual quality and quantitative metrics.

The present invention can effectively improve the visual quality of low-light and low-resolution facial images and solve the problem of loss of important facial information in existing cascade technology.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of the present invention, wherein IC-FSRNet is mainly used to restore the face structure and realize structure-fidelity reconstruction, while DENet can further enhance facial details and improve visual quality.

FIG2 is a schematic diagram of the overall structure of IC-FSRNet of the present invention.

Figure 3 is a comparison of the subjective results of the present invention with several other SOTA methods, (a): I _LLR subjective results; (b): subjective results of the SCTANet+LLformer method; (c): subjective results of the LLformer+SCTANet method; (d): subjective results of the SISN+FECNet method; (e): subjective results of the FECNet+SISN method; (f): subjective results of the SFMNet+LEDNet method; (g): subjective results of the LEDNet+SFMNet method; (h): subjective results of the IC-FSRNet method; (i): subjective results of the IC-FSRDENet method (j): corresponding high-quality face image reference standards.

FIG4 is a cosine similarity comparison diagram of the present invention and several other SOTA methods.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

In order to improve the quality of low-light low-resolution face images, a natural idea is to design a joint low-light low-resolution face super-resolution framework. However, simple cascade schemes (LLIE+FSR or FSR+LLIE) are defective in capturing complex facial features, resulting in unsatisfactory super-resolution results with obvious chromatic aberration and structural loss. To solve this dilemma, the present invention proposes to decompose this task into structural fidelity reconstruction and texture consistency learning. The former is tailored for improving the quality of low-resolution low-light face images while maintaining structural fidelity, while the latter focuses on eliminating perturbations and artifacts caused by low-light degradation and reconstruction.

Considering that the brightness-adjusted region can provide complementary information to improve the super-resolution of face images, and conversely, the super-resolution face images can also promote the brightness adjustment of face images. Therefore, the present invention introduces interactive learning to utilize the brightness-adjusted region and the reconstructed information, and realizes iterative refinement. In view of this, IC-FSRNet constructs a brightness estimation (IE) branch for estimating brightness adjustment coefficients and a face super-resolution (FSR) branch that combines these coefficients to jointly perform super-resolution and brightness adjustment. IC-FSRNet uses an iterative approach to enhance face reconstruction by fully exploring the mutual enhancement between super-resolution and brightness adjustment. After maintaining structural fidelity, the present invention utilizes the advantages of the diffusion model in synthesizing fine image details and designs a detail enhancement network (DENet) based on the diffusion probability model to eliminate artifacts and improve visual quality. The method is generally referred to as IC-FSRDENet.

The goal of the present invention is to generate high-quality super-resolution face images from a given low-light low-resolution face image I _LLR . Existing methods usually focus on a single task, resolution upscaling or brightness correction, but rarely consider both degradations at the same time. Therefore, they cannot effectively handle this complex task. Although cascade solutions (i.e., performing face super-resolution first and then brightness adjustment (FSR+LLIE) or performing brightness adjustment on the image first and then super-resolution on the image (LLIE+FSR)) have been introduced to restore these degraded face images, the cascade method ignores the intrinsic relationship between the two tasks and has limited performance improvement. The face images generated by the cascade scheme usually lack face structure and details, and there is obvious color deviation. To alleviate this problem, the present invention proposes a novel method to decompose the task into structural fidelity reconstruction and texture consistency learning. The former is designed to improve the quality of I _LLR faces while maintaining structural fidelity. The latter focuses on eliminating perturbations and artifacts caused by low-light degradation and reconstruction. As shown in Figure 1, a super-resolution reconstruction method for face images under low-light conditions, the reconstruction method comprises the following steps:

Step 1: Synthesize a low-light low-resolution face image I _LLR ;

Step 2: Construct the brightness correction face super-resolution network IC-FSRNet;

Step 3: Input the image synthesized in step 1 into step 2 to improve the brightness of the face image and restore the face structure information to obtain I _SR1 ;

Step 4: Build the detail enhancement model DENet;

Step 5: Input the image obtained in step 3 into step 4 to improve the facial details of the face image, so that the face image has a better visual effect, and obtain I _SR2 .

First, the I _LLR face image is input into IC-FSRNet for brightness adjustment, and a roughly restored super-resolution face image is generated at the same time; considering that the improved features of brightness adjustment and face super-resolution are complementary, the super-resolution face image can improve the brightness adjustment, and the brightness-adjusted area can provide complementary information to promote face super-resolution; in view of this, IC-FSRNet includes two branches, a brightness estimation branch (IE) for estimating the brightness adjustment coefficient, and a super-resolution branch to combine these coefficients to jointly perform super-resolution and brightness adjustment; in addition, IC-FSRNet iteratively performs brightness estimation and face super-resolution, exploring the mutual information between them to promote face super-resolution; although IC-FSRNet can well restore the face structure and complete the brightness adjustment, the restored face image always lacks high-frequency details and has noise caused by low-light degradation and reconstruction; therefore, a detail enhancement network DENet based on a diffusion model is further constructed, and the results restored by IC-FSRNet are input into it to improve face details and visual quality. The method is collectively called IC-FSRDENet.

Furthermore, the brightness correction face super-resolution network constructed specifically includes a brightness estimation IE branch for estimating the brightness coefficient for brightness adjustment and a face super-resolution FSR branch for reconstructing the super-resolution result; considering that the improved features of brightness adjustment and face super-resolution are complementary, the super-resolution face image can improve the brightness adjustment, and the brightness-adjusted area can provide supplementary information to promote face super-resolution, so IC-FSRNet iteratively performs brightness coefficient estimation and super-resolution to explore the complementary relationship between the two for mutual refinement;

First, the low-light low-resolution face image I _LLR is sent to two branches;

Then, the brightness correction super-resolution block ICSRB takes the output features of the previous step as input to encode their mutual relations;

ICSRB first performs face super-resolution and brightness estimation in parallel; the former is done by the commonly used RCAB (Imagesuper-resolution using very deep residual channel attention networks) to generate enhanced features S ₁ ; the latter is achieved through efficient bilateral grid learning. Thanks to bilateral grid learning, the present invention can estimate the brightness coefficient grid in the low-resolution dimension and then slice it into the high-resolution space with a guide map, which is more efficient.

Furthermore, the brightness correction super-resolution block ICSRB converts the output features of the previous step As input to encode their mutual relationship, specifically, firstly three cascaded convolutional layers are used from/> Predict the bilateral brightness grid B ₁ ; at the same time, extract the guidance map G from the upsampled I _LLR through the convolution layer; with B ₁ and G, the brightness coefficient C ₁ is obtained through 3D slicing, expressed as,

C ₁ = f _Slice (B ₁ ,G),

Where f _Slice represents a 3D slice operation;

Furthermore, the 3D slice first projects the brightness grid B ₁ onto a 3D grid using G, where the first two dimensions of the grid represent the 2D position in the image plane and the third dimension represents the image intensity of G; then the grid is blurred by Gaussian blur; finally, based on the blurred bilateral grid and the guide image G, the brightness coefficient C ₁ is obtained by accessing the grid value using trilinear interpolation.

The brightness coefficient _C1 is then fed into the brightness adjustment block to adjust the illumination of feature _S1 in the FSR branch, producing the adjusted result Given that more accurate brightness coefficients can facilitate face super-resolution, while higher-resolution face images can improve brightness estimation, ICSRB achieves mutual enhancement between brightness estimation and face super-resolution in an iterative manner.

Furthermore, in addition to using the brightness coefficient from the IE branch to improve the brightness in the FSR branch, the brightness adjustment result is also Feedback to the brightness refinement block of the IE branch to refine the original brightness and improve the subsequent brightness estimation; in this way, face super-resolution can promote brightness estimation; then, the generated super-resolution result and brightness refinement result are input into the following L-1 ICSRBs for more accurate mutual refinement and utilization;

After L times of brightness estimation and face super-resolution, a convolutional layer is applied to the result of the FSR branch to generate the reconstruction result I _SR1 ; in order to optimize the network, L1 pixel loss is used,

Among them, I _HR is the reference standard for high-quality face images.

Furthermore, the brightness adjustment is specifically that the present invention uses a conversion function with a learnable brightness coefficient to perform brightness adjustment. Specifically, first, the learned coefficient _Ci is mapped to [-1,1] through a convolution layer and a Sigmoid function to obtain a brightness coefficient _C'i . The super-resolution feature _Si will be input into the cascaded convolution layer to generate _S'i , and then the brightness adjustment is performed through the following conversion function:

in represents the brightness-adjusted feature, and * indicates pixel-wise multiplication.

Furthermore, the brightness refinement is specifically as follows: considering that face super-resolution and brightness estimation can promote each other, the present invention feeds back the super-resolution result of the face super-resolution branch to the brightness estimation branch to refine the brightness estimation; considering the global characteristics of brightness information, a cross-attention mechanism is introduced to explore global information from the super-resolution results; specifically, given the super-resolution result and the original brightness _Bi , using different fully connected layers to generate query Q from _Bi , and from/> Get the key K and value V; then calculate the cross attention between them,

Where F _Att represents the attention map, f _Softmax represents softmax, and d is a hyperparameter;

Then, a feed-forward network is applied to generate the refined result; in this way, the super-resolution features can be utilized to refine the brightness estimation in the next step.

Furthermore, the construction of the detail enhancement model DENet is specifically as follows: although IC-FSRNet can improve the brightness of the face image and restore the face structure, the face image restored by IC-FSRNet lacks face relationship details and has obvious artifacts, and cannot provide a pleasant visual experience. For this reason, a detail enhancement network (DENet) is established to improve the visual quality. Inspired by the powerful generation ability of the diffusion probability model DDPM, the present invention designs a DENet based on conditional DDPM, and uses I _SR1 as additional auxiliary information to help reverse the diffusion process. Specifically, the training objectives of DENet are as follows,

Where ∈ ～ N (0, 1), (x, y) corresponds to I _SR1 and I _HR , γ ～ p (γ), and f _DENet represents DENet. Thanks to DENet, the method of the present invention can restore visually pleasing face images.

Verification test

In order to study low-light low-resolution face super-resolution and verify the effectiveness of the proposed method, low-light low-resolution face and corresponding high-quality face image datasets are essential. However, there is currently no paired low-light low-resolution face dataset, and it is particularly difficult to construct a real face image pair. Therefore, the present invention simulates the degradation process and uses the existing face dataset to synthesize low-light low-resolution face images and normal light high-resolution face image pairs.

The key parts of the degradation simulation of low-light low-resolution face images include brightness adjustment and noise addition. The present invention describes the degradation simulation in detail below.

Brightness adjustment. The purpose of brightness adjustment is to transform a normal face image into a low-light face image. The present invention does not directly use gamma correction, but uses a combination of linear transformation and gamma transformation to perform brightness adjustment, which can better approximate low-light images. Specifically, given a normal-light face image I _HR , the brightness adjustment can be expressed as: I _LL = β×(α×I _HR ) ^γ , where β∈U(0.5,1), α∈U(0.9,1), γ∈U(1.5,5), and I _LL represents a low-light face image.

Noise addition. In addition to changes in illumination, low-light environments also introduce noise. Therefore, the present invention adds noise to low-light face images. In simulating real low-light noise, the present invention takes into account the characteristics and effects of in-camera image signal processing (ISP). ISP refers to the processing steps performed by camera hardware and software to convert raw sensor data into the final image. When simulating low-light noise, it is also important to consider the characteristics of ISP. Therefore, the noise addition process can be expressed as: Where _fISP and/> represents the ISP function and the inverse ISP function, _NP and _NG correspond to Poisson noise and Gaussian noise, and _ILLN is the generated low-light noisy face image. Here, the ISP function consists of white balance, demosaicing, CAM-to-XYZ, XYZ-to-RGB, and tone mapping.

Downsampling. Finally, the present invention uses Bicubic to downsample I _LLN by a factor of ×16 to generate a low-light and low-resolution face image I _LLR .

Datasets and Metrics The model of the present invention is trained on CelebAMaskHQ. The present invention randomly selects 3050 face images from the data as the training set, and uses an additional 300 face images as the test set. First, the present invention crops the face images to 256×256 as high-resolution face images, and then downsamples them to 16×16 as I _LLR images using the above-mentioned degradation process. The present invention selects peak signal-to-noise ratio (PSNR), structural similarity (SSIM), learned perceptual image patch similarity (LPIPS), and natural image quality evaluator (NIQE) indicators as evaluation indicators.

Experimental details: The number L of ICSRBs is 14. IC-FSRNet and DENet are trained separately. IC-FSRNet is trained first, and then the parameters of IC-FSRNet are fixed. DENet is trained using the output of IC-FSRNet as auxiliary information. Adam is selected as the optimizer of the model of the present invention, and the learning rate is set to 1e-4 throughout the training stage.

Comparison Method In the experiment of the present invention, since there is no low-light low-resolution face super-resolution method, the present invention selects several representative face super-resolution methods and image enhancement methods, cascades them, and forms 18 comparison methods. The face super-resolution methods include SISN (Face Hallucination via Split-Attention in Split-Attention Network), SCTANet (A Spatial Attention-Guided CNN-Transformer Aggregation Network for Deep Face Image Super-Resolution), SFMNet (Spatial-Frequency Mutual Learning for Face Super-Resolution), and the low-light image enhancement methods include FECNet (Deep Fourier-Based Exposure Correction Network with Spatial-Frequency Interaction), LLformer (Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method) and LEDNet (Lednet: Joint low-light enhancement and deblurring in the dark).

Comparative Test

Subjective Results Figure 3 shows the super-resolution results of several face images selected from the test set. It can be seen that the simple cascade solution will produce artifacts and cannot effectively restore the face structure. The reason is that the individual methods only focus on a single degradation, while the input suffers from both low-light and low-resolution degradation. That is, the former process will inevitably produce artifacts and errors, which will interfere with the latter process and cause error accumulation, ultimately leading to the loss of important face information. In contrast, the method of the present invention takes into account low-light and low-resolution degradations and proposes a joint framework for restoring I _LLR face images. In terms of visual quality, IC-FSRNet successfully restores clear face structures that the cascade solution cannot restore. However, the faces reconstructed by IC-FSRNet are still noisy and lack fine details. Thanks to DENet, the method of the present invention is able to restore visually pleasing high-quality face images.

Table 1. Objective comparison of our method with several other SOTA methods. The best results are marked in bold and the suboptimal results are underlined.

Objective Results Table 1 shows the objective performance of the comparison methods and the proposed method. Specifically, the upper part of Table 1 gives the results of the method that performs face super-resolution first and then performs low-light image enhancement, while the middle part lists the results of the method that performs low-light image enhancement first and then performs face super-resolution. The bottom shows the performance of the proposed method. From the comparison in Table 1, it can be seen that it is difficult for a simple cascade solution to achieve satisfactory performance, while the joint method of the present invention is significantly better than the cascade solution. In particular, compared with the cascade solution, IC-FSRNet shows the best performance in the three evaluation indicators except NIQE, especially in PSNR and SSIM. Compared with the second best performing method, it improves PSNR by 3.04 dB and SSIM by 0.0661 while using fewer parameters. The addition of DENet further improves LPIPS and NIQE, but affects PSNR and SSIM to a certain extent. Despite this, the PSNR and SSIM of IC-FSRDENet are still significantly higher than those of other methods. In general, the results in Table 1 confirm the superiority of the proposed method over the simple cascade solution.

Comparison of face recognition results In addition to restoring high-quality face images, face super-resolution methods should also improve the performance of downstream tasks such as face recognition. Therefore, the present invention compares and analyzes the method proposed in the present invention and the existing cascade method in terms of face recognition performance. In order to evaluate the face recognition performance, the pre-trained face recognition model Deepface is used to extract face identity features from face images reconstructed by different methods and the corresponding high-quality face image reference standards. Subsequently, the cosine similarity between these face identity features is calculated as a measure of face recognition performance. The comparison results are shown in Figure 4. As shown in the figure, the cosine similarity of the method of the present invention is significantly higher than that of the comparison method. This result clearly shows that the method of the present invention is superior to other methods in terms of face recognition performance, highlighting the effectiveness and superiority of the method proposed in the present invention in improving downstream tasks of face recognition.

Considering that brightness adjustment and face super-resolution can provide complementary information and promote each other, IC-FSRNet constructs a brightness estimation branch for estimating brightness adjustment coefficients and a face super-resolution branch for face super-resolution, and iteratively performs brightness coefficient estimation and face super-resolution, allowing them to complement each other and promote face reconstruction. Afterwards, DENet based on the diffusion probability model is used to remove artifacts and improve the quality of rough results. Experimental results show that the method of the present invention achieves state-of-the-art performance.

Claims (8)

1. A method for super-resolution reconstruction of facial images under low light conditions, characterized in that the reconstruction method comprises the following steps: Step 1: Synthesize a low-light low-resolution face image I _LLR ; Step 2: Construct the brightness correction face super-resolution network IC-FSRNet; Step 3: Input the image synthesized in step 1 into step 2 to improve the brightness of the face image and restore the face structure information to obtain I _SR1 ; Step 4: Build the detail enhancement model DENet; Step 5: Input the image obtained in step 3 into step 4 to improve the facial details of the face image, so that the face image has a better visual effect, and obtain I _SR2 . 2. According to the method for super-resolution reconstruction of facial images under low-light conditions in claim 1, it is characterized in that the construction of the brightness correction face super-resolution network specifically includes a brightness estimation IE branch for estimating the brightness coefficient for brightness adjustment and a face super-resolution FSR branch for reconstructing the super-resolution result; First, the low-light low-resolution face image I _LLR is sent to two branches; Then, the brightness correction super-resolution block ICSRB takes the output features of the previous step as input to encode their mutual relations. 3. According to the method for super-resolution reconstruction of face images under low light conditions in claim 2, it is characterized in that the brightness correction super-resolution block ICSRB converts the output feature Specifically, three cascaded convolutional layers are used as input to encode their mutual relationship. Predict the bilateral brightness grid B ₁ ; at the same time, extract the guidance map G from the upsampled I _LLR through the convolution layer; with B ₁ and G, the brightness coefficient C ₁ is obtained through 3D slicing, expressed as, C ₁ = f _Slice (B ₁ ,G), Where f _Slice represents a 3D slice operation; The brightness coefficient _C1 is then fed into the brightness adjustment block to adjust the brightness of the feature _S1 in the FSR branch, producing the adjusted result 4. According to the method for super-resolution reconstruction of facial images under low-light conditions in claim 3, it is characterized in that the 3D slice first uses G to project the brightness grid _B1 onto the 3D grid, The mesh is then blurred using Gaussian blur; Finally, according to the blurred bilateral grid and the guidance image G, the brightness coefficient C ₁ is obtained by accessing the grid values using trilinear interpolation. 5. According to claim 3, a method for super-resolution reconstruction of face images under low light conditions is characterized in that in addition to using the brightness coefficient from the IE branch to improve the brightness in the FSR branch, the brightness adjusted result is also Feedback to the brightness refinement block of the IE branch to refine the original brightness and improve the subsequent brightness estimation; in this way, face super-resolution can promote brightness estimation; then, the generated super-resolution result and brightness refinement result are input into the following L-1 ICSRBs for more accurate mutual refinement and utilization; After L times of brightness estimation and face super-resolution, a convolutional layer is applied to the result of the FSR branch to generate the reconstruction result I _SR1 ; in order to optimize the network, L1 pixel loss is used, Among them, I _HR is the reference standard for high-quality face images. 6. According to claim 5, a method for super-resolution reconstruction of facial images under low-light conditions, characterized in that the brightness adjustment is specifically to first map the learned coefficient _Ci to [-1,1] through a convolution layer and a Sigmoid function to obtain a brightness coefficient _C'i . The super-resolution feature _Si will be input into the cascaded convolution layer to generate _S'i , and then brightness adjustment is performed through the following conversion function: in represents the brightness-adjusted feature, and * indicates pixel-wise multiplication. 7. According to the method for super-resolution reconstruction of face images under low light conditions in claim 6, the brightness refinement is specifically: given the super-resolution result and the original brightness _Bi , using different fully connected layers to generate query Q from _Bi , and from/> Get the key K and value V; then calculate the cross attention between them, Where F _Att represents the attention map, f _Softmax represents softmax, and d is a hyperparameter; Then, a feed-forward network is applied to generate the refined result; the super-resolution features are used to refine the brightness estimation in the next step. 8. According to the method for super-resolution reconstruction of face images under low-light conditions in claim 2, it is characterized in that the construction of the detail enhancement model DENet is specifically that the training objective of DENet is as follows: Where ∈～N(0,1), (x,y) corresponds to I _SR1 and I _HR , γ～p(γ), and f _DENet represents DENet.