A linear low-light image enhancement method and system based on knowledge distillation

Abstract

The application discloses a linear low-light image enhancement method and system based on knowledge distillation, and comprises the following steps: obtaining a high-order curve parameter map by using a teacher network for a low-light image training sample and an exposure condition map, obtaining a teacher network enhancement result, and optimizing and training the teacher network by taking a no-reference loss function as a target; obtaining a linear enhancement curve parameter map by using a student network for the low-light image training sample and the exposure condition map, enhancing the low-light image training sample according to the linear enhancement curve parameter map, obtaining a student network enhancement result, and optimizing and training the student network by taking a loss between the student network enhancement result and the teacher network enhancement result as a target; and obtaining a linear enhancement curve parameter map by using the trained student network for a low-light image to be enhanced and an exposure condition map, so as to obtain an enhancement result of the low-light image to be enhanced. The low-light image enhancement is light, low in power consumption and high in performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a linear faint light image enhancement method and system based on knowledge distillation. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] For faint light image enhancement, existing solutions are mainly divided into traditional methods and deep learning-based methods. Early traditional methods widely adopt histogram equalization technology and Retinex theory. Histogram equalization considers that the histogram of an ideal image is approximately uniformly distributed, and the distribution of the histogram is stretched to enhance the contrast of the image. The Retinex theory separates the illumination component and the incident component in the image through logarithmic operation, transforms them into the frequency domain and enhances them using different strategies, and then converts them back to the image space and obtains the final result through exponential operation. Based on the Retinex theory, the image is decomposed into reflection and illumination, and it is assumed that the reflection component is consistent under any lighting condition, so the illumination enhancement problem is converted into an illumination estimation problem.

[0004] However, due to the complexity of the algorithm and the accuracy of the imaging model, traditional faint light image enhancement methods do not work well when dealing with diverse real scenes, and are prone to amplify noise, exacerbate color cast, and introduce artifacts.

[0005] The initial deep learning-based faint light image enhancement method mostly uses simulated paired data sets for supervised training. With the rapid development of generative adversarial networks and reference-free learning, some faint light image enhancement methods that are not based on paired data have been proposed. For example, the paper "Zero-reference deep curve estimation for low-light image enhancement" published in CVPR in 2020 proposed the Zero-DCE algorithm, which no longer considers the faint light image enhancement problem as an image-to-image transformation problem, but redefines it as an image-specific curve estimation problem, and combines a reference-free loss function to realize a lightweight faint light image enhancement network. Although Zero-DCE greatly reduces the computational burden of the network, it still cannot meet the demand for low-power lightweight faint light enhancement methods for embedded systems or end-side devices with limited computing power. SUMMARY

[0006] To solve the above problems, the application provides a linear low-light image enhancement method and system based on knowledge distillation, which analyzes the high-order curve mapping process in the teacher network, solves the redundant calculation process by approximating the mapping relationship at a certain place of the high-order curve to the tangent line at the place through the knowledge distillation method, replaces the high-order exponential operation with linear calculation, takes the teacher network enhancement result as supervision, trains a lightweight and low-power student network, and realizes low-light image enhancement with lightweight, low power consumption and high performance.

[0007] To achieve the above purpose, the application adopts the following technical solutions:

[0008] In the first aspect, the application provides a linear low-light image enhancement method based on knowledge distillation, which includes:

[0009] The teacher network is used to obtain a high-order curve parameter map from the low-light image training sample and the exposure condition map, the low-light image training sample is iteratively mapped according to the high-order curve parameter map, the teacher network enhancement result is obtained, and the teacher network is optimized and trained with a no-reference loss function as the target;

[0010] The student network is used to obtain a linear enhancement curve parameter map from the low-light image training sample and the exposure condition map, the low-light image training sample is visually enhanced according to the linear enhancement curve parameter map, the student network enhancement result is obtained, and the student network is optimized and trained with the loss between the student network enhancement result and the teacher network enhancement result as the target;

[0011] The trained student network is used to obtain a linear enhancement curve parameter map from the low-light image to be enhanced and the exposure condition map, so as to obtain the enhancement result of the low-light image to be enhanced.

[0012] As an optional implementation, the teacher network is used to obtain a high-order curve parameter map from the exposure condition map and the low-light image training sample, which is consistent in size with the low-light image training sample, after eight convolutional layers, three convolutional and RuLU activation function iterative operations in each convolutional layer, and one convolutional and Tanh activation function.

[0013] As an optional implementation, the process of iteratively mapping the low-light image training sample according to the high-order curve parameter map includes: dividing the three high-order curve parameter maps into n groups as a group, corresponding to n times of iterative mapping, in each iterative mapping, the three high-order curve parameter maps act on the RGB three channels of the low-light image respectively, and after the mapping is completed, a high-order brightness enhancement curve, i.e. the teacher network enhancement result, is obtained.

[0014] As an optional implementation, in the n th mapping: is the n th group of high-order curve parameter maps; I is the input low-light image, LE n-1(I) is the mapping result of the last time.

[0015] As an alternative embodiment, the no-reference loss function comprises a self-supervised spatial exposure control loss, a spatial consistency loss, a color constancy loss and an illumination smoothness loss.

[0016] As an alternative embodiment, the mapping process of the pixel-wise high-order curve parameter map in the student network adopts a tangent line for approximation.

[0017] As an alternative embodiment, the student network is used to train the exposure condition map and the low-light image training sample, and a 4-fold downsampling is first performed, then seven convolution layers are passed through, and after the iteration operation of three convolutions and RuLU activation functions in each convolution layer, a 4-fold upsampling is performed to obtain a linear enhancement curve parameter map.

[0018] The linear enhancement curve parameter and the low-light image are brought into the linear enhancement curve for pixel-wise operation to obtain a student network enhancement result. k is a slope, B is a bias parameter, and I is an input low-light image.

[0019] In a second aspect, the present application provides a linear low-light image enhancement system based on knowledge distillation, comprising:

[0020] The teacher network training module is configured to obtain a high-order curve parameter map from a teacher network for the low-light image training sample and the exposure condition map, perform iterative mapping on the low-light image training sample according to the high-order curve parameter map to obtain a teacher network enhancement result, and optimize and train the teacher network with a no-reference loss function as a target.

[0021] The student network distillation module is configured to obtain a linear enhancement curve parameter map from a student network for the low-light image training sample and the exposure condition map, perform visual enhancement on the low-light image training sample according to the linear enhancement curve parameter map to obtain a student network enhancement result, and optimize and train the student network with a loss between the student network enhancement result and the teacher network enhancement result as a target.

[0022] The enhancement module is configured to obtain a linear enhancement curve parameter map from the trained student network for the low-light image to be enhanced and the exposure condition map, so as to obtain an enhancement result of the low-light image to be enhanced.

[0023] In a third aspect, the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein when the computer instructions are run by the processor, the method of the first aspect is completed.

[0024] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, complete the method of the first aspect.

[0025] Compared with the prior art, the present application has the following beneficial effects:

[0026] The present application proposes a linear low-light image enhancement method and system based on knowledge distillation, which improves the existing method of using deep learning technology to estimate the curve parameters to realize low-light image enhancement. First, the teacher network is trained based on the Zero-DCE method, and the no-reference loss function is adjusted according to the exposure condition map during the training process to establish the connection between the exposure condition map and the optimization target, thereby realizing the controllable low-light image enhancement teacher network. Second, through the analysis of the high-order curve mapping process in the teacher network, it is found that there is a redundant calculation process, and then the mapping relationship at a certain place of the high-order curve is approximated to the tangent line at that place through the knowledge distillation method, and the linear calculation is used instead of the complex high-order exponential operation, thereby greatly reducing the calculation amount and greatly improving the calculation efficiency of the algorithm. The knowledge distillation method is used to realize the lightweight, low-power and high-performance low-light image enhancement.

[0027] The present application proposes a linear low-light image enhancement method and system based on knowledge distillation, which designs a lightweight student network. The student network takes the down-sampled low-light image as input, and takes the enhanced result generated by the high-order curve mapping of the teacher network as supervision to distill the high-order curve mapping knowledge, and outputs the slope and bias parameters required for approximate linear change. After up-sampling to the original size, the parameters are used to enhance the visual quality of the low-light image.

[0028] The network model of the present application is a low-power lightweight model. Only about 3K parameters, the calculation amount and power consumption are much smaller than those of other existing networks, and the inference speed is significantly improved.

[0029] Unlike the existing image-to-image conversion method, the method of the present application is obtained by linear adjustment of the input image, which will not produce too many artifacts, thereby greatly improving the detection performance, and is friendly to computer vision advanced tasks such as target detection, and the enhancement result can significantly improve the performance of advanced application algorithms.

[0030] Unlike the existing algorithm which can only output a single enhancement result, the present application can produce different enhancement results by recommending or customizing the exposure condition map with one model, and the enhancement result is personalized and controllable, which meets the personalized needs of users and application scenarios. In addition, the no-reference loss function can be combined with the exposure condition as input to realize controllable exposure adjustment.

[0031] Advantages of the additional aspects of the present application will become apparent in light of the following description. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated by reference herein. The embodiments depicted are provided to explain the present application and are not intended to limit the application thereto.

[0033] Figure 1 The basic flowchart of embodiment 1 of the present application is provided as follows;

[0034] Figure 2 (a)- Figure 2 (o) The enhanced results of the embodiment method of the present application and other existing methods are compared as shown in the following diagram. DETAILED DESCRIPTION

[0035] The present application will be further described with reference to the accompanying drawings and examples.

[0036] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0037] It should be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise. In addition, it should be understood that the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a list of steps or units is not necessarily limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or devices.

[0038] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0039] Embodiment 1

[0040] The present embodiment provides a linear low-light image enhancement method based on knowledge distillation, as shown in the following formula (1), which includes the following steps: Figure 1

[0041] The teacher network is used to obtain a high-order curve parameter map from the low-light image training sample and the exposure condition map, and the low-light image training sample is iteratively mapped according to the high-order curve parameter map to obtain a teacher network enhancement result, and the teacher network is optimized and trained with a no-reference loss function as the target.​

[0042] The linear enhancement curve parameter map is obtained by using the student network on the low-light image training sample and the exposure condition map, the low-light image training sample is visually enhanced according to the linear enhancement curve parameter map, and the student network enhancement result is obtained, and the student network is optimized and trained with the loss between the student network enhancement result and the teacher network enhancement result as the target;

[0043] The linear enhancement curve parameter map is obtained by using the trained student network on the low-light image to be enhanced and the exposure condition map, so as to obtain the enhancement result of the low-light image to be enhanced.

[0044] In this embodiment, the process of training the teacher network is to train a teacher network with large parameter quantity and strong performance by using the pixel-by-pixel high-order brightness enhancement curve and the no-reference loss function proposed by Zero-DCE;

[0045] Specifically, it includes:

[0046] (1) taking the exposure condition map and the low-light image as input;

[0047] Wherein, in the training stage, an arbitrary shape region is randomly selected from the exposure condition map, and different random values (ranging from 0.2 to 0.8) are assigned to the inside and outside of the region;

[0048] (2) taking the default parameters of the high-order brightness enhancement curve as output;

[0049] Wherein, the exposure condition map and the low-light image are input into the teacher network, the structure of the teacher network is shown in Table 1, after 8 convolution layers, after 3 times of Conv convolution in each convolution layer, the iteration operation of RuLU activation function is connected, and then one convolution and Tanh activation function are connected, and finally the high-order curve parameter map with the same size as the input low-light image is estimated;

[0050] The high-order curve parameter map is divided into n groups as a group, corresponding to n times of iteration mapping, in each iteration mapping, the three parameter maps are respectively applied to the RGB three channels of the low-light image, and after the mapping is completed, the high-order brightness enhancement curve specific to the low-light image is obtained;

[0051] Wherein, in the first mapping: is the normalized low-light image, is the first group of high-order curve parameter maps;

[0052] In the n-th mapping: is the n-th group of high-order curve parameter maps;

[0053] The final teacher network enhancement result is R = LE n (I).

[0054] Table 1 Network layers of the teacher network and the student network

[0055]

[0056] (3) In the training process, a no-reference loss function L tea is used as the optimization target to complete the training of the teacher network, and after the training is completed, the parameters of the teacher network are fixed and the next stage is entered;

[0057] The no-reference loss function L tea is:

[0058] L tea = λ sec L sec + λ sc L sc + λ cc L cc + λ is L is

[0059] Wherein, L sec is the self-supervised spatial exposure control loss; L sc is the spatial consistency loss; L cc is the color constancy loss; L is is the illumination smoothness loss; λ sec , λ sc , λ cc , λ is are the coefficients of each loss, which can be set to 10, 1, 5, 200 respectively.

[0060] Specifically: (3-1) Self-supervised spatial exposure control loss L sec can measure the distance between the average intensity value of the local area of the teacher network enhanced result R and the exposure condition graph E; therefore, the exposure condition graph can control the exposure level of the final result:

[0061]

[0062] Wherein, M is the number of non-overlapping local areas m with a size of 16x16, and Mean(·) is the average intensity value in the area.

[0063] (3-2) Spatial consistency loss L sc is to ensure the spatial consistency between the input low-light image and its adjusted low-light image, which is represented as:

[0064]

[0065] Wherein, K is the number of 4x4 local areas i, and Ω(i) is the four adjacent areas (up, down, left, right) centered on area i.

[0066] (3-3) Color constancy loss L cc ; according to the gray world color constancy assumption, that is, the color in each sensor channel is averaged to be gray on the whole image; color constancy loss L cc Potential color deviation in the final enhancement result can be corrected, and the relationship between the three adjusted color channels is represented as:

[0067]

[0068] where (p, q) e e = {(r, g), (r, b), (g, b)} represents a pair of channels.

[0069] (3-4) Illumination smoothing loss L is to maintain the monotonic relationship between adjacent pixels in each curve parameter map , which is defined as:

[0070]

[0071] where N is the number of iterations, and and represent horizontal and vertical gradient operations, respectively.

[0072] In this embodiment, the process of training the student network is to use the teacher network enhancement result generated by the trained teacher network as supervision to train a lightweight, low-power student network; the student network approximates the calculation process of the high-order brightness enhancement curve pixel by pixel with its tangent, thereby saving a large number of iterative operation processes.

[0073] Specifically, it includes:

[0074] (1) Take the exposure condition map and the low-light image as input;

[0075] The network structure of the student network is shown in Table 1; first, 4 times down-sampling is performed, and after passing through 7 convolution layers, 3 times Conv convolution is performed in each convolution layer, followed by iteration operation of the RuLU activation function, and then 4 times up-sampling is performed to obtain a linear enhancement curve default parameter map with the same size as the input low-light image; different from the teacher network, the student network adopts a U-shaped network structure, which has a faster inference speed.

[0076] (2) Take the linear enhancement curve default parameter as output, and bring the parameter and the low-light image into the linear enhancement curve for pixel-by-pixel operation to generate the student network enhancement result as k is the slope and B is the bias parameter.

[0077] (3) In the process of distilling the student network, the parameters of the teacher network are frozen, and the teacher network enhancement result is used as supervision. The L1 loss between the student network enhancement result and the teacher network enhancement result is used as the target to supervise and train the student network.

[0078] The L1 loss is:

[0079]

[0080] In actual application, only the lightweight and low-power student network is used to process the faint light image. According to different scenes, the exposure condition map recommended in the embodiment or the exposure condition map customized by the user according to personal needs can be used to generate faint light enhancement images with different enhancement effects.

[0081] Specifically, the exposure condition map recommended in the embodiment assigns a large exposure value to the underexposed area and a small exposure value to the well-exposed / overexposed area.

[0082] First, the brightness channel L of the input faint light image and its average value L avg are obtained. avg Then, the spatially varying exposure map is calculated by S+A*Norm(L

[0083] For underexposed images and overexposed images, A is set to 0.15 by experience, and S is set to 0.55 and 0.25, respectively. This spatial variation setting can achieve satisfactory performance for fine exposure adjustment.

[0084] The selected exposure condition map and faint light image are input into the trained student network to obtain the parameters of the linear enhancement curve and The parameters and the faint light image are input into the linear enhancement curve for pixel-by-pixel operation to generate the enhancement result.

[0085] In the embodiment, 1000 normal brightness images are selected from the REDS dataset for training. For the underexposure correction experiment, 219 pairs of paired data from the VELOL-L-Cap dataset are used. Some underexposed image datasets are used for experiments, including NPE, LIME, MEF, DICM, and VV. The test set contains various underexposed images, denoted as UndExp-Web.

[0086] Figure 2 (a)- Figure 2(o) as shown, the method of the present embodiment is compared with 11 methods (15 models), including eight underexposure enhancement methods (SCI (CVPR'22), URetinex-Net (CVPR'22), RUAS (CVPR'21), Zhao et al. (ICCV'21), EnlightenGAN (TIP'21), ZeroDCE (CVPR'20), DRBN (CVPR'20), Retinex-Net (BMVC'18)) and three exposure correction methods (Afifi et al. (CVPR'21), ExCNet (ACMMM'19), Exposure (TOG'18)).

[0087] Comparison is made using pre-trained models, including two models of RUAS (RUAS-5K and RUAS-LOL) and three models of SCI (SCI-easy, SCI-medium and SCI-difficult). Among these methods, SCI, RUAS, Zero-DCE, ExCNet and the method of the present embodiment are zero-reference learning based methods, while the remaining methods require supervised training. Due to the limited robustness of RUAS-LOL, the visual results of RUAS-LOL are not shown.

[0088] Since the method of the present embodiment can quickly deduce multiple results by setting exposure condition maps with different exposure values, the present embodiment provides two versions of results. "Ours-auto": the present embodiment fixes the exposure values of the exposure condition maps for all test images, uses a uniform exposure value of 0.65 for underexposed images, and sets 0.2 for overexposed images; "Ours-manu": the present embodiment manually sets the exposure values of the exposure condition maps for each test image, generates results using only three fixed exposure values (0.55, 0.65 and 0.75 for underexposed images; 0.2, 0.25 and 0.3 for overexposed images), and subjectively selects the result that is visually most pleasing as the final result. When selecting the final result, no reference image can be accessed. The results using "Ours-auto" are used for visual comparison.

[0089] Embodiment 2

[0090] The present embodiment provides a linear low-light image enhancement system based on knowledge distillation, comprising:

[0091] The teacher network training module is configured to obtain a high-order curve parameter map from the teacher network for the low-light image training sample and the exposure condition map, iteratively map the low-light image training sample according to the high-order curve parameter map to obtain a teacher network enhancement result, and optimize and train the teacher network with a no-reference loss function as the target.

[0092] The student network distillation module is configured to obtain a linear enhancement curve parameter map from the micro-light image training sample and the exposure condition map by using the student network, perform visual enhancement on the micro-light image training sample according to the linear enhancement curve parameter map, obtain a student network enhancement result, and optimize and train the student network by taking the loss between the student network enhancement result and the teacher network enhancement result as a target.

[0093] The enhancement module is configured to obtain a linear enhancement curve parameter map from the to-be-enhanced micro-light image and the exposure condition map by using the trained student network, and obtain an enhancement result of the to-be-enhanced micro-light image.

[0094] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules can be executed in a computer system such as a set of computer executable instructions as part of the system.

[0095] In more embodiments, there are also provided:

[0096] An electronic device includes a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method described in Embodiment 1 is completed. For brevity, it will not be repeated here.

[0097] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0098] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, and a part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0099] A computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the method described in Embodiment 1 is completed.

[0100] The method in the embodiment 1 can be directly embodied by a hardware processor or by a combination of hardware and software modules in the processor. The software modules can be located in a storage medium in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or the like. The storage medium is located in a memory, and a processor reads information in the memory and combines hardware to complete the steps of the above method. To avoid repetition, no further detailed description is given herein.

[0101] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0102] Although the specific embodiments of the present application are described above in combination with the drawings, the description is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A linear low-light image enhancement method based on knowledge distillation, characterized in that, include: To enhance low-light images and exposure condition maps, a linear enhancement curve parameter map is obtained using a trained student network. Based on the obtained linear enhancement curve parameter map, pixel-by-pixel operations are performed on the low-light image to be enhanced, yielding the enhancement result. in, The trained student network is a knowledge distillation of the trained teacher network. This knowledge distillation approximates the higher-order curve parameter map obtained by the trained teacher network as a linear augmentation curve parameter map. The student network is trained in the following manner: For low-light image training samples and exposure condition maps, a linear enhancement curve parameter map is obtained using a student network. Based on the linear enhancement curve parameter map, visual enhancement is performed on the low-light image training samples to obtain the student network enhancement result. The student network is trained with the loss between the student network enhancement result and the teacher network enhancement result as the objective, so that the high-order curve parameter map obtained by using the trained teacher network is approximately the linear enhancement curve parameter map. The teacher network is trained in the following manner: For low-light image training samples and exposure condition maps, a teacher network is used to obtain a high-order curve parameter map. The low-light image training samples are iteratively mapped according to the high-order curve parameter map to obtain the teacher network enhancement result. The teacher network is then optimized and trained with a no-reference loss function as the objective.

2. The linear low-light image enhancement method based on knowledge distillation as described in claim 1, characterized in that, The results of the teacher network enhancement were obtained in the following manner: For the low-light image training samples and exposure condition map, a high-order curve parameter map is obtained using the trained teacher network. The low-light image training samples are then iteratively mapped based on the high-order curve parameter map to obtain the teacher network enhancement result.

3. A linear low-light image enhancement method based on knowledge distillation as described in claim 1 or 2, characterized in that, The teacher network consists of eight convolutional layers, one convolution and Tanh activation function, wherein each convolutional layer undergoes three iterations of convolution and ReLU activation function.

4. A linear low-light image enhancement method based on knowledge distillation as described in claim 1 or 2, characterized in that, The process of iteratively mapping the low-light image training samples based on the higher-order curve parameter map includes: The high-order curve parameter maps are divided into n groups of three, corresponding to n iterative mappings. In each iterative mapping, the three high-order curve parameter maps are applied to the three RGB channels of the low-light image respectively. After the mapping is completed, a high-order brightness enhancement curve is obtained. This high-order brightness enhancement curve is used by teachers when enhancing low-light images using a network.

5. The linear low-light image enhancement method based on knowledge distillation as described in claim 4, characterized in that, During the nth mapping: ) This is the parameter map of the nth group of higher-order curves, where I is the input low-light image. This is the result of the previous mapping.

6. The linear low-light image enhancement method based on knowledge distillation as described in claim 1, characterized in that, The no-reference loss function includes: a self-supervised spatial exposure control loss to characterize the distance between the average intensity value of a local region in the teacher network augmentation result and the exposure condition map; a spatial consistency loss to characterize the spatial consistency between the teacher network augmentation result and the low-light image training samples; a color constancy loss to correct potential color deviations in the teacher augmentation result and establish the relationship between color channels; and an illumination smoothing loss to maintain the monotonic relationship between adjacent pixels in the higher-order curve parameter map.

7. The linear low-light image enhancement method based on knowledge distillation as described in claim 6, characterized in that, The self-supervised spatial exposure control loss is determined by the following formula: in, For self-supervised spatial exposure control loss, M is the number of non-overlapping regions m, Mean(·) is the average intensity value within the non-overlapping regions, R is the teacher network enhancement result, and E is the exposure condition map.

8. The linear low-light image enhancement method based on knowledge distillation as described in claim 7, characterized in that, The spatial consistency loss is determined by the following formula: in, K represents the spatial consistency loss, K is the number of local regions i, and Ω(i) is the top, bottom, left, and right regions centered on local region i.

9. The linear low-light image enhancement method based on knowledge distillation as described in claim 7, characterized in that, The color constancy loss is determined by the following formula: in, For color constancy loss, (p, q) ∈ ε = {(r, g), (r, b), (g, b)} represents a pair of channels.

10. The linear low-light image enhancement method based on knowledge distillation as described in claim 6, characterized in that, The illumination smoothing loss is determined by the following formula: Where N is the number of iterations, and ξ = {r, g, b}, and These represent the horizontal and vertical gradient operations, respectively, and A is the curve parameter graph.

11. The linear low-light image enhancement method based on knowledge distillation as described in claim 1, characterized in that, The method of approximating the high-order curve parameter map obtained by using the trained teacher network as a linear enhancement curve parameter map includes: approximating the pixel-by-pixel mapping process of the high-order curve parameter map in the student network using tangents.

12. The linear low-light image enhancement method based on knowledge distillation as described in claim 1, characterized in that, The student network consists of: 4x downsampling, seven convolutional layers, and 4x upsampling. Specifically, each convolutional layer undergoes three iterations of convolution and ReLU activation; the pixel-by-pixel operation on the low-light image to be enhanced, based on the obtained linear enhancement curve parameter map, includes: By substituting the linear enhancement curve parameters and the low-light image into the linear enhancement curve and performing pixel-by-pixel operations, the student network enhancement result is obtained. k is the slope, B is the bias parameter, and I is the input low-light image.

13. The linear low-light image enhancement method based on knowledge distillation as described in claim 1, characterized in that, The exposure condition map assigns large exposure values ​​to underexposed areas and small exposure values ​​to well-exposed or overexposed areas.

14. The linear low-light image enhancement method based on knowledge distillation as described in claim 13, characterized in that, The exposure condition map is obtained in the following manner: Obtain the brightness channel L and its average value Lavg of the input low-light image. The spatial variation exposure map is calculated by S+A×Norm(Lavg-L), where S is the base exposure value, A is the adjustment range, and the Norm operation normalizes its input to the range of [-1, 1]. For underexposed and overexposed images, set A to 0.15 and S to 0.55 and 0.25 respectively.

15. A linear low-light image enhancement system based on knowledge distillation, characterized in that, include: The enhancement module is configured to use the low-light image to be enhanced and the exposure condition map, and a trained student network to obtain a linear enhancement curve parameter map. Based on the obtained linear enhancement curve parameter map, pixel-by-pixel operations are performed on the low-light image to be enhanced to obtain the enhancement result. in, The trained student network is a knowledge distillation of the trained teacher network. This knowledge distillation makes the high-order curve parameter map obtained by the trained teacher network approximate as a linear augmentation curve parameter map. The student network is trained in the following manner: For low-light image training samples and exposure condition maps, a linear enhancement curve parameter map is obtained using a student network. Based on the linear enhancement curve parameter map, visual enhancement is performed on the low-light image training samples to obtain the student network enhancement result. The student network is trained with the loss between the student network enhancement result and the teacher network enhancement result as the objective, so that the high-order curve parameter map obtained by using the trained teacher network is approximately the linear enhancement curve parameter map. The teacher network is trained in the following manner: For low-light image training samples and exposure condition maps, a teacher network is used to obtain a high-order curve parameter map. The low-light image training samples are iteratively mapped according to the high-order curve parameter map to obtain the teacher network enhancement result. The teacher network is then optimized and trained with a no-reference loss function as the objective.

16. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1 to 14.

Images