A fast low-illumination image enhancement method and system based on PairLIE and meta-learning

Abstract

The application relates to a fast low-illumination image enhancement method and system based on PairLIE and meta-learning, and aims to enable a low-illumination image enhancement model to learn initial model parameters with good universality and migration ability from multiple low-illumination tasks. When the model faces new light conditions or new scenes that are not involved in training, only a small amount of low-illumination images and a small number of parameter updates are needed to quickly adapt to the degradation characteristics of the current low-illumination images, thereby reducing the time and computing resources required for large-scale retraining for new scenes. Meanwhile, the application uses the self-supervised learning mechanism of PairLIE based on paired low-illumination images to decompose and constrain the illumination component and the reflection component of the image, so that the model training can be completed without relying on strictly paired normal-illumination reference images, so as to reduce the acquisition and labeling costs of normal-illumination reference data and reduce the dependence of the model on large-scale paired training data.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, digital image processing and deep learning, and in particular to a fast low-light image enhancement method and system based on PairLIE and meta-learning. Background Technology

[0002] Low-light images typically suffer from insufficient overall brightness, uneven exposure in local areas, low contrast, loss of image details, and significant noise. These issues not only affect human visual perception but also reduce the processing efficiency of subsequent visual tasks such as object detection, object recognition, video surveillance, industrial inspection, and medical image analysis. Currently, low-light image enhancement methods mainly include traditional image processing methods, supervised learning methods based on deep learning, and unsupervised or zero-reference learning methods.

[0003] Traditional image processing methods mainly include histogram equalization and methods based on Retinex theory. Histogram equalization improves image contrast by adjusting the pixel distribution of grayscale or color channels. Typical methods include global histogram equalization and contrast-limited adaptive histogram equalization (CLAHE). These methods are simple to implement and computationally fast, but they are prone to amplifying image noise and may cause problems such as over-enhancement in local areas, color distortion, and unnatural image brightness.

[0004] Retinex-based methods typically represent low-light images as the product of the reflectance and illumination components. Image brightness enhancement is achieved by estimating and adjusting the illumination component, such as single-scale Retinex, multi-scale Retinex, and multi-scale Retinex with color restoration. These methods can preserve image structure and color information to some extent, but their enhancement effect is highly dependent on the accuracy of the illumination component estimation. When images contain complex lighting, strong light-dark boundaries, or non-uniform illumination, phenomena such as halos, edge artifacts, local overexposure, and loss of detail are prone to occur.

[0005] With the development of deep learning technology, researchers have begun to use convolutional neural networks to establish mapping relationships between low-light and normal-light images. For example, MBLLEN, Retinex-Net, and end-to-end low-light imaging methods typically use pairs of low-light and normal-light images for supervised training. These methods can learn the recovery rules of image brightness, color, and texture from training data, usually achieving better enhancement results than traditional methods. However, high-quality pairs of low-light and normal-light images are difficult to acquire, and the two images need to maintain strict consistency in terms of scene, object position, and camera parameters, leading to high data acquisition and annotation costs. Furthermore, when the lighting conditions, imaging equipment, or scene distribution of the actual image differ significantly from the training data, the model's enhancement performance can easily decline significantly.

[0006] To reduce reliance on paired normal illumination images, existing technologies have proposed unsupervised learning and zero-reference learning methods. For example, EnlightenGAN utilizes generative adversarial networks to learn the distribution relationship between low-light and normal illumination images; Zero-DCE transforms low-light image enhancement into an image-specific curve estimation problem; and SCI employs lightweight networks to improve enhancement speed and model flexibility. These methods do not require strictly paired normal illumination reference images, reducing the difficulty of obtaining training data, but their model parameters are usually still obtained under a fixed training data distribution. When faced with new lighting scenes that differ significantly from the training environment, problems such as noise amplification, color shift, uneven brightness, or local over-enhancement may still occur.

[0007] The PairLIE low-light image enhancement method is based on Retinex theory. It uses paired low-light images from the same scene for self-supervised learning and obtains intermediate image representations, illumination components, and reflection components through prediction networks, illumination estimation networks, and reflection estimation networks, respectively. This method leverages the reflection consistency between paired low-light images to learn the image enhancement model without using a normal illumination reference image, thereby reducing the dependence on paired labeled data.

[0008] However, PairLIE and other existing low-light image enhancement models typically employ fixed network initialization parameters and training strategies, with their training primarily targeting pre-defined data distributions. When ambient light intensity, light source type, image exposure, scene content, or camera sensor changes in real-world applications, the degradation distribution of low-light images also alters. Especially when there are significant differences in lighting between different scenes or when lighting conditions change dynamically, existing models struggle to quickly adapt to the new image distribution, potentially leading to significant fluctuations in enhancement results. To achieve better enhancement effects, it is usually necessary to retrain or fine-tune the model using new scene data for an extended period, which is detrimental to the rapid deployment of the model in dynamic monitoring, mobile terminals, and real-time vision systems.

[0009] Furthermore, meta-learning methods, such as Model-Independent Meta-Learning (MAML), can learn highly generalized initial model parameters across multiple tasks, enabling the model to adapt to new tasks with only a small number of samples and gradient updates. However, existing low-light image enhancement methods typically focus on learning the mapping relationship from low-light images to enhanced images on a fixed dataset, and have not yet fully addressed the issues of rapid transfer and parameter adaptation between different lighting tasks.

[0010] In summary, existing low-light image enhancement techniques have at least the following shortcomings: First, they typically treat the low-light degradation process as a relatively fixed process, making it difficult to adapt to the complex, random, and dynamically changing lighting conditions in real-world environments. Second, supervised learning methods rely on a large amount of paired training data, resulting in high data acquisition costs. Third, while unsupervised or zero-reference methods reduce data requirements, their generalization ability across scenes, devices, and large-scale lighting variations remains limited. Fourth, when models face new low-light environments, they typically require a large amount of training data and a long retraining or fine-tuning time, making it difficult to achieve rapid and stable scene adaptation. Summary of the Invention

[0011] This invention provides a fast low-light image enhancement method and system based on PairLIE and meta-learning to solve the problems of existing low-light image enhancement methods, which are usually based on fixed training data distribution for model training. When faced with changes in ambient light intensity, light source type, scene content, exposure degree and imaging device, they are prone to insufficient generalization ability, unstable enhancement effect and need to be retrained or fine-tuned for a long time using new scene data.

[0012] In a first aspect, the present invention provides a fast low-light image enhancement method based on PairLIE and meta-learning, comprising: Different low-light images are constructed into multiple low-light image enhancement tasks. Each task samples from the low-light task distribution p(T), denoted as: {T i}i =1 ~p(T), where T i Let p(T) represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-dimensional training process. For each low-light image enhancement task, in-task learning is performed in the PairLIE network to obtain the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are... i It is used for inter-task meta-learning, and the common initial parameter θ0 is optimized across tasks to obtain the meta-training parameter θ_T; After meta-training, when a new low-light scene that was not included in the training appears, the meta-training parameter θ_T is used as the initial parameter of the model. Only a small number of low-light images of the new scene are needed and a few parameter updates are performed to obtain an enhanced model suitable for the new scene and perform image enhancement on the low-light images of the new scene. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure or illumination distribution.

[0013] Furthermore, the in-task learning part of PairLIE includes a pair of low-light image input units, a projection network P-Net, an illumination estimation network L-Net, a reflection estimation network R-Net, a loss calculation unit, and an in-task parameter update unit; The parameters of P-Net, L-Net, and R-Net together constitute the parameters to be learned: θ = {θ_P, θ_L, θ_R} Here, θ_P, θ_L, and θ_R represent the network parameters of P-Net, L-Net, and R-Net, respectively. The networks of the same type in the two image processing branches have the same network structure and share the corresponding parameters to ensure that the paired low-light images are processed according to the same projection and decomposition rules.

[0014] Furthermore, P-Net, L-Net, and R-Net are trained together using projection loss, reflection consistency loss, and Retinex reconstruction loss; the three types of losses are used to constrain the fidelity of projection information, the consistency of paired image reflection components, and the reconstructibility of illumination-reflection decomposition, respectively.

[0015] Furthermore, within a specific low-light task, the task-specific parameter θ is learned based on the PairLIE total loss. i For the i-th low-light enhancement task, the PairLIE network is first initialized with the common initial parameter θ0: θ i (0) = θ0 Then, the total in-task loss is calculated using paired low-light images from the task, and one or more gradient updates are performed: θ i (s+ ¹ ) = θ i (s) α _{θ i (s)} _total(T i ;θ i (s) ) Where α represents the in-task learning rate, and s represents the number of in-task updates; after a preset number of updates K, the task-specific parameter θ corresponding to the i-th task is obtained. i =θ i (K) ; Different lighting tasks yielded θ1, θ2, ..., θ i Task-specific parameters reflect the model's parameter state after minor updates to the current lighting conditions.

[0016] Furthermore, the inter-task meta-learning part optimizes the common initial parameter θ0 in reverse based on the adaptation loss after a small number of updates for multiple tasks, so that the common initial parameter can converge quickly with a small number of gradient updates when facing different low-light tasks. The meta-loss is obtained by summing or averaging the query set losses of multiple tasks in a meta-training batch. _meta = (1 / N) Σ i=1 _total(Q i ;θ i ) Update the common initial parameters based on the meta-loss: θ0← θ0 β _{θ0} _meta Where β represents the step size of the inter-task meta-update; when using the full MAML update method, the gradient of the meta-loss is backpropagated to the common initial parameters through the intra-task parameter update process; the inter-task meta-parameter update uses the Adam optimizer, which repeatedly executes task sampling, intra-task update, query set loss calculation and inter-task meta-update until the meta-loss converges or reaches the preset number of training rounds, to obtain the final meta-training parameters θ_T.

[0017] Furthermore, after meta-training is complete, when the model encounters a new low-light scene T_new that did not participate in meta-training, it first loads the meta-training parameters as the initial parameters for the new task: θ_new (0) = θ_T Subsequently, a small number of paired low-light images from the new scene are acquired, and a few sub-gradient updates are performed using the same projection loss, reflection consistency loss, and Retinex reconstruction loss as in the meta-training phase: θ_new (s+ ¹ ) = θ_new (s) α_adapt _{θ_new (s)} _total(T_new) After a limited number of updates, the parameter θ_new suitable for the new low-light scene is obtained.

[0018] Furthermore, for the low-light image I_low to be enhanced, it is first input into P-Net to obtain the intermediate projected image i: i = P(I_low;θ_P) Then, the intermediate projected image is input into L-Net and R-Net respectively to obtain the illumination component L and the reflection component R: L = L(i; θ_L), R = R(i; θ_R) Nonlinear correction of the illumination component: L′ = L^λ Where λ is the illumination correction factor; for illumination maps with pixel values ​​normalized to the range of [0, 1], when 0 < λ < 1, the value of the illumination component can be increased, thereby brightening low-illuminance areas. The corrected illumination component is multiplied pixel by pixel with the reflection component to obtain the enhanced image: I_enhanced = R ⊙ L′ = R ⊙ L^λ Finally, the range of pixel values ​​in the enhanced image is limited: I_out = clip(I_enhanced, 0, 1) Converting I_out to an 8-bit image, a 16-bit image, or a floating-point image yields an enhanced result that improves brightness while preserving color, texture, and structural information.

[0019] Secondly, the present invention provides a fast low-light image enhancement system based on PairLIE and meta-learning, comprising: The task construction module is used to construct multiple low-light image enhancement tasks from different low-light images. Each task is sampled from the low-light task distribution p(T), denoted as: {T i} i =1 ~p(T), where T i Let p(T) represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-dimensional training process. The task learning module is used to perform in-task learning within the PairLIE network for each low-light image enhancement task, obtaining the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are... i It is used for inter-task meta-learning, and the common initial parameter θ0 is optimized across tasks to obtain the meta-training parameter θ_T; The image enhancement module is used to enhance the image of a new low-light scene that was not included in the training when a new low-light scene appears after meta-training. The meta-training parameter θ_T is used as the initial parameter of the model. Only a small number of low-light images of the new scene are needed and a few parameter updates are performed to obtain an enhancement model suitable for the new scene and enhance the low-light images of the new scene. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure or illumination distribution.

[0020] Thirdly, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the fast low-light image enhancement method based on PairLIE and meta-learning as described above.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the fast low-light image enhancement method based on PairLIE and meta-learning as described above.

[0022] The above-described technical solution of the present invention has the following advantages: The fast low-light image enhancement method based on PairLIE and meta-learning provided in the first aspect of this invention does not train a set of model parameters for only one fixed lighting condition, but instead learns universal meta-training parameters θ_T from multiple low-light enhancement tasks. When the model encounters a new scene, exposure condition, light source type, or imaging device, it uses the meta-training parameters θ_T as initial parameters and only needs to perform a limited number of gradient updates using a small number of low-light images of the new scene to obtain task-specific parameters suitable for the current scene. Therefore, this method can reduce the time and computational cost required for the model to be trained from scratch or fine-tuned for a long time for new scenes, and improve the adaptation speed of the low-light image enhancement model.

[0023] This method enables low-light image enhancement models to learn initial parameters with good generality and transferability from various low-light tasks. When the model faces new lighting conditions or scenes not used in training, it only needs a small number of low-light images and a few parameter updates to quickly adapt to the degradation features of the current low-light images, thereby reducing the time and computational resources required for large-scale retraining for new scenes. Simultaneously, this method utilizes PairLIE, a self-supervised learning mechanism based on paired low-light images. By decomposing and constraining the illumination and reflection components of the images, it completes model training without relying on strictly paired normal-light reference images, reducing the cost of collecting and labeling normal-light reference data and decreasing the model's dependence on large-scale paired training data.

[0024] It is understood that the beneficial effects of the second, third, and fourth aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description

[0025] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 The schematic diagram illustrates the principle of the fast low-light image enhancement method based on PairLIE and meta-learning provided by this invention. Detailed Implementation

[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0028] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0029] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a particular feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized. "A plurality" means "two or more."

[0031] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] Specifically, this invention addresses the following technical problems: Existing low-light image enhancement models treat illumination degradation as a relatively fixed process, making it difficult to adapt to complex, random, and dynamically changing illumination conditions in real-world environments; existing models require a large number of training samples and long periods of retraining or fine-tuning when facing new low-light scenes, exposure conditions, or imaging devices; when there are distributional differences between different training data and actual application data, the model enhancement effect is prone to significant fluctuations; existing methods are prone to uneven brightness, local overexposure, color shift, noise amplification, and loss of image details in complex illumination environments; supervised low-light image enhancement methods rely on a large number of rigorously registered low-light-normal-light image pairs, resulting in high data acquisition and production costs; the invention aims to improve the model's generalization ability, adaptation speed, and enhancement stability across scenes, exposure conditions, and imaging devices, making it more suitable for dynamic monitoring, mobile terminals, industrial vision, and other practical application scenarios requiring rapid deployment.

[0033] Therefore, the core technical problem to be solved by this invention is: how to enable a low-light image enhancement model to obtain a universal initialization capability for different lighting tasks when there is a lack of a large amount of normal illumination pairing data, and how to enable the model to achieve fast and stable image enhancement through a small number of samples and a limited number of parameter updates when encountering a new low-light environment.

[0034] 1. Overall Working Principle This invention first constructs multiple low-light image enhancement tasks from low-light images obtained under different scenes, exposure levels, light source conditions, or imaging devices. Each task samples from the low-light task distribution p(T), represented as: {T i} i =1 ~p(T) Among them, T i Let represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-order training process. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure, or illumination distributions.

[0035] For each low-light image enhancement task, first in the attached... Figure 1 Intra-task learning is performed in the PairLIE network on the left to obtain the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are... i For attach Figure 1The inter-task meta-learning on the right optimizes the common initial parameter θ0 across tasks to obtain the meta-training parameter θ_T. After meta-training, when a new low-light scene that was not included in the training appears, the meta-training parameter θ_T is used as the initial parameter of the model. Only a small number of low-light images of the new scene are needed, and a few parameter updates are performed to obtain an enhanced model suitable for the new scene.

[0036] 2. Structure and connectivity of the learning component within the PairLIE task As attached Figure 1 As shown on the left, the in-task learning part of PairLIE includes a pairwise low-light image input unit, a projection network P-Net, an illumination estimation network L-Net, a reflection estimation network R-Net, a loss calculation unit, and an in-task parameter update unit. The parameters of P-Net, L-Net, and R-Net together constitute the parameters to be learned. θ = {θ_P, θ_L, θ_R} Here, θ_P, θ_L, and θ_R represent the network parameters of P-Net, L-Net, and R-Net, respectively. Networks of the same type in the two image processing branches have the same network structure and share corresponding parameters to ensure that paired low-light images are processed according to the same projection and decomposition rules.

[0037] 2.1 Paired Low-Light Image Input From the i-th low-light enhancement task T i Two low-light images, I1 and I2, from the same scene are selected. These two images can be low-light images taken under different exposure times, different lighting intensities, or different light source conditions. According to Retinex theory, the two images are represented as follows: I1 = R ⊙ L1 I² = R ⊙ L² Here, R represents the scene reflection component, which mainly includes object color, texture, outline, and structural information; L1 and L2 represent the illumination components corresponding to the two images; and ⊙ represents pixel-wise multiplication. Since the two images come from the same scene, their reflection components should be the same or approximately the same, while their illumination components can be different. If there are slight changes in viewpoint, camera shake, or object movement between the two images, registration, cropping, scale unification, and effective region selection can be performed before inputting them into the network.

[0038] 2.2 P-Net Projection Processing As attached Figure 1 As shown, the first low-light image I1 is input into the first P-Net branch to obtain the first intermediate projection image i1; the second low-light image I2 is input into the second P-Net branch to obtain the second intermediate projection image i2. I1 = P(I1;θ_P) I2 = P(I2;θ_P) Although two P-Net bounding boxes are drawn separately in the attached figure, the two P-Nets share the parameter θ_P. The P-Net is used to perform feature projection on the original low-light image, preserving the main structural and texture information in the image, while reducing sensor noise, random noise, and interfering features that are not suitable for direct Retinex decomposition.

[0039] 2.3 Estimation of Illumination and Reflection Components The first intermediate projection image i1 is input into L-Net and R-Net respectively to obtain the first illumination component L1 and the first reflection component R1; the second intermediate projection image i2 is input into the same L-Net and R-Net respectively to obtain the second illumination component L2 and the second reflection component R2. L1= L(i1; θ_L), R1= R(i1; θ_R) L2= L(i2; θ_L), R2= R(i2; θ_R) The L-Net in both processing branches shares the parameter θ_L, and the R-Net in both processing branches shares the parameter θ_R. The L-Net is used to extract brightness distribution information that varies with ambient lighting, exposure level, and light source conditions; the R-Net is used to extract color, texture, edge, and scene structure information that is relatively unchanging with lighting.

[0040] 3. In-task loss calculation for PairLIE As attached Figure 1 As shown by the dashed lines on the left, this invention trains P-Net, L-Net, and R-Net together using projection loss, reflection consistency loss, and Retinex reconstruction loss. These three types of losses are used to constrain projection information fidelity, pairwise image reflection component consistency, and the reconstructability of illumination-reflection decomposition, respectively.

[0041] 3.1 Projection Loss The projection loss is used to constrain the difference between the intermediate projected image output by the P-Net and the original input image, and its expression is: _P = ∥I1 i1∥2² This loss prevents P-Net from destroying the main scene content of the original image when removing noise and interfering features. In an extended implementation, the projection loss can be calculated for both branches simultaneously: _P = ∥I1 i1∥2² + ∥I2 i2∥2² The projection loss is marked in the attached figure using the first image processing branch as an example, but this does not mean that the projection loss can only be calculated in the first branch.

[0042] 3.2 Loss of Reflection Consistency Since the input images I1 and I2 originate from the same scene, their reflection components should be consistent. Therefore, the reflection consistency loss is calculated based on the first reflection component R1 and the second reflection component R2: _C = ∥R1 R2∥2² Reflection consistency loss corresponding to the appendix Figure 1 The lower left side is connected by R1 and R2. The dashed lines connecting _C(R1, R2) represent the relationship between images. Through this loss, the network attributes the differences between pairs of low-light images primarily to the illumination component, rather than the reflection component, thus improving the stability of color, texture, and structural information extraction. Simultaneously, since the random noise in the two images is typically different, restricting the consistency of the reflection component also reduces the network's fitting to the random noise of a single image.

[0043] 3.3 Retinex Reconstruction Loss Construct the Retinex reconstruction loss based on the first intermediate projection image i1, the first illumination component L1, and the first reflection component R1: _R = ∥R1⊙ L1 i1∥2² + ∥R1 i1 / (stopgrad(L1)+ε)∥2² + ∥L1 L0∥2² + ∥ L1∥1 Where L0 is the initial illumination estimate obtained based on the maximum values ​​of the red, green, and blue channels of each pixel in the input image: L0(x) = maxI1 c (x), c ∈ {R, G, B} Where stopgrad(·) means to stop the gradient propagation operation; ε is a positive number set to prevent division by zero; L1 represents the spatial gradient of the illumination component. The first term ensures that the estimated illumination and reflection components can be reconstructed into an intermediate projected image; the second term uses the illumination component to guide the estimation of the reflection component; the third term constrains the difference between the predicted illumination map and the initial illumination estimate; and the fourth term keeps the illumination map spatially smooth, reducing residual texture and noise in the illumination map.

[0044] In another embodiment, the second Retinex reconstruction loss can also be calculated based on i2, L2, and R2 of the second branch, and the reconstruction losses of the two branches can be added together or averaged.

[0045] 3.4 Total Losses During the Mission By weighting and combining the above three types of losses, we obtain the total loss within the PairLIE task: _total = w_P _P + w_C _C + w_R _R Where w_P, w_C, and w_R are the weights of the projection loss, reflection consistency loss, and Retinex reconstruction loss, respectively. In a specific embodiment, w_P=500, w_C=0.5, and w_R=0.5 can be set. The above values ​​are only preferred settings and can be adjusted according to the image noise level, exposure range, and training stability.

[0046] 4. Intrinsic parameters θ i The acquisition Appendix Figure 1 The "Learn θ" text shown at the top left i by "_total in PairLIE" means: within a specific low-light task, the task-specific parameter θ is learned based on the total loss of PairLIE. i For the i-th low-light enhancement task, the PairLIE network is first initialized with the common initial parameter θ0: θ i (0) = θ0 Then, the total in-task loss is calculated using paired low-light images from the task, and one or more gradient updates are performed: θ i (s+ ¹ ) = θ i (s) α _{θ i (s)} _total(T i ;θ i (s) ) Where α represents the in-task learning rate, and s represents the number of in-task updates; after a preset number of updates K, the task-specific parameter θ corresponding to the i-th task is obtained. i =θ i (K); Different lighting tasks yielded θ1, θ2, ..., θ i Task-specific parameters reflect the model's parameter state after minor updates to the current lighting conditions.

[0047] Preferably, each task is further divided into a support set S i and query set Q i The support set is used to calculate the in-task loss and obtain the task-specific parameter θ. i The query set is used to evaluate the adaptation effect of task-specific parameters and for subsequent meta-parameter updates between tasks. When the number of original images for a single scene is small, paired images can be cropped into multiple non-overlapping image patches, which constitute the support set and query set respectively.

[0048] 5. Inter-task learning section As attached Figure 1 As shown on the right, the horizontal arrows indicate that the common initial parameters are gradually optimized from θ0 to the meta-training parameters θ_T, and the vertical boxes indicate θ1, θ2, ..., θ_T. i This represents the task-specific parameters learned through PairLIE within different low-light tasks. The meta-learning module does not simply select parameters for a single task, but rather back-optimizes the common initial parameter θ0 based on the adaptation loss obtained after a small number of updates across multiple tasks. This ensures that the common initial parameter can converge quickly with a small number of gradient updates when facing different low-light tasks.

[0049] The meta-loss is obtained by summing or averaging the query set losses of multiple tasks in a meta-training batch. _meta = (1 / N) Σ i=1 _total(Q i ;θ i ) Update the common initial parameters based on the meta-loss: θ0← θ0 β _{θ0} _meta Here, β represents the step size of the inter-task meta-update. When using the full MAML update method, the gradient of the meta-loss is backpropagated to the common initial parameters through the intra-task parameter update process. The Adam optimizer can be used for inter-task meta-parameter updates. By repeatedly executing task sampling, intra-task updates, query set loss calculation, and inter-task meta-updates until the meta-loss converges or reaches the preset number of training rounds, the final meta-training parameters θ_T are obtained.

[0050] θ0──{θ1, θ2,…,θ i}──→θ_T The above parameter relationship means that θ0 is used as the common initial parameter for different tasks. The initial parameter is then meta-optimized by the task-specific parameters formed by multiple tasks and their adaptation loss, and finally the parameter θ_T with the ability to adapt quickly across tasks is obtained.

[0051] 6. Rapid adaptation to new low-light scenes After meta-training is complete, when the model encounters a new low-light scene T_new that did not participate in meta-training, it first loads the meta-training parameters as the initial parameters for the new task: θ_new (0) = θ_T Subsequently, a small number of paired low-light images from the new scene are acquired, and a few sub-gradient updates are performed using the same projection loss, reflection consistency loss, and Retinex reconstruction loss as in the meta-training phase: θ_new (s+ ¹ ) = θ_new (s) α_adapt _{θ_new (s)} _total(T_new) After a limited number of updates, the parameters θ_new suitable for the new low-light scene are obtained. Since θ_T is a parameter jointly learned on multiple different low-light tasks, it contains common knowledge across different illumination degradation tasks. Therefore, the adaptation for the new scene does not need to be retrained from random parameters. The new task adaptation phase does not require a normal illumination reference image; only a small number of low-light images or pairs of low-light images are needed. For situations with high real-time requirements or where there are no new scene adaptation samples, the meta-trained parameters θ_T can also be used directly for image enhancement.

[0052] 7. Enhanced Image Generation Appendix Figure 1 This mainly illustrates the training and parameter optimization process, with the image enhancement output process implemented after the illustrated network. For the low-light image I_low to be enhanced, it is first input into the P-Net to obtain the intermediate projected image i: i = P(I_low;θ_P) Then, the intermediate projected image is input into L-Net and R-Net respectively to obtain the illumination component L and the reflection component R: L = L(i; θ_L), R = R(i; θ_R) Nonlinear correction of the illumination component: L′ = L^λ Where λ is the illumination correction factor. For an illumination map with pixel values ​​normalized to the range [0, 1], when 0 < λ < 1, the value of the illumination component can be increased, thus brightening low-light areas. The corrected illumination component is then multiplied pixel-by-pixel with the reflection component to obtain the enhanced image: I_enhanced = R ⊙ L′ = R ⊙ L^λ Finally, the range of pixel values ​​in the enhanced image is limited: I_out = clip(I_enhanced, 0, 1) Depending on the application requirements, I_out can be converted to an 8-bit image, a 16-bit image, or a floating-point image to obtain an enhanced result that improves brightness while preserving color, texture, and structural information.

[0053] Key points of this invention: 1. A two-layer training framework combining in-task learning and meta-learning optimization between tasks in PairLIE Construct a two-layer optimization structure that includes intra-task learning and inter-task meta-learning: During the in-task learning process, the PairLIE network is used to process paired low-light images in a specific low-light task, and the corresponding task-specific parameters θ are obtained based on the total in-task loss. i ; In the inter-task meta-learning process, the common initial parameter θ0 is meta-optimized by integrating the task-specific parameters and their adaptation losses corresponding to multiple low-light tasks, so as to obtain the meta-training parameter θ_T with the ability to quickly adapt across tasks.

[0054] This two-layer training framework enables the model to learn "how to quickly adapt to different lighting tasks" instead of just learning a fixed low-light data distribution.

[0055] 2. A method for constructing meta-learning tasks based on different low-light degradation conditions Based on at least one of the following factors—scene, exposure level, ambient light intensity, light source type, light source color temperature, noise level, or imaging device—low-light image data is divided into multiple low-light image enhancement tasks.

[0056] Each task includes at least two low-light images with different lighting or exposure conditions in the same scene, and can be further divided into a support set for intra-task updates and a query set for inter-task meta-updates.

[0057] By constructing the tasks described above, different low-light degradation conditions are treated as independent but related learning tasks to participate in meta-training, thereby enabling the model to acquire common knowledge among different low-light environments.

[0058] 3. Self-supervised in-task learning mechanism based on the reflectance consistency of paired low-light images For a pair of low-light images I1 and I2 in the same scene, intermediate projection images i1 and i2, illumination components L1 and L2, and reflection components R1 and R2 are obtained by using P-Net, L-Net, and R-Net with shared parameters, respectively.

[0059] Taking advantage of the fact that the reflectance properties of the same scene remain basically unchanged under different lighting conditions, a reflectance consistency constraint is imposed on R1 and R2, which enables the model to attribute the changes between paired images mainly to the changes in lighting, and to learn a low-light image enhancement model without using a normal illumination reference image.

[0060] 4. A method for generating multiple task-specific parameters from common initial parameters and then optimizing the meta-parameters in reverse. For multiple low-light tasks, starting from a common initial parameter θ0, gradient updates are performed once or multiple times within the task to obtain θ1, θ2, ..., θ i Then, based on the loss adapted to each task, a meta-loss is constructed, and the gradient of the meta-loss is passed to the common initial parameters through the intra-task update process. The common initial parameters are then updated, and finally the meta-training parameters θ_T are obtained.

[0061] The key is not to use a specific optimizer, but to use the in-task adaptation results of multiple low-light enhancement tasks to back-optimize the common initialization parameters of the low-light image enhancement network, so that it can quickly approach the applicable parameters for the new task with a few updates.

[0062] 5. A rapid adaptation method with few samples for new low-light environments When the model faces a new low-light scene that was not involved in the meta-training, the augmentation network is initialized with the meta-training parameters θ_T, a small number of low-light images or pairs of low-light images in the new scene are obtained, and the parameters are updated a limited number of times using the same or corresponding self-supervised loss as in the meta-training stage to obtain the new scene adaptation parameters.

[0063] This process does not require retraining from random parameters or providing a reference image with normal illumination, thus enabling rapid deployment for new scenes, new exposure conditions, new light source types, or new imaging devices.

[0064] Compared with existing low-light image enhancement methods, this invention combines the PairLIE image decomposition network with a meta-learning mechanism. Through intra-task learning and inter-task meta-learning, it obtains model parameters that can quickly adapt to different low-light conditions, and has the following beneficial effects: 1. Able to quickly adapt to new low-light scenarios Corresponding to the intra-task parameter update and inter-task meta-parameter update processes in this invention, this invention does not train a set of model parameters for only one fixed lighting condition, but learns universal meta-training parameters θ_T from multiple low-light enhancement tasks.

[0065] When the model encounters new scenes, exposure conditions, light source types, or imaging devices, it uses the meta-training parameter θ_T as initial parameters and only needs to perform a limited number of gradient updates using a small number of low-light images of the new scene to obtain task-specific parameters suitable for the current scene. Therefore, this invention can reduce the time and computational cost required for the model to be trained from scratch or fine-tuned for a long time for new scenes, and improve the adaptation speed of low-light image enhancement models.

[0066] 2. Improve the model's generalization ability and enhance its stability under different lighting conditions. Corresponding to the technical content of meta-training based on multiple low-light tasks in this invention, this invention constructs multiple low-light image enhancement tasks according to scene, exposure level, light intensity, light source type, noise level or imaging device, and uses multiple tasks to jointly optimize the common initial parameters of the model.

[0067] The resulting meta-training parameters contain common information across different low-light degradation conditions, reducing the model's dependence on a single training data distribution. Even when real-world lighting conditions differ from the training data, the model can still perform image enhancement relatively stably, mitigating issues such as brightness fluctuations, local overexposure, color shifts, and unstable detail recovery caused by scene changes. Experiments also demonstrate that this approach exhibits good cross-scene adaptability on the SICE dataset, which includes various complex exposure conditions.

[0068] 3. Reduce reliance on paired images under normal illumination and manually labeled data. Corresponding to the PairLIE task learning part of this invention, this invention uses paired low-light images with different lighting or exposure levels in the same scene for training, without needing to provide a strictly registered normal-light reference image for each low-light image.

[0069] This invention utilizes the reflectance consistency between paired low-light images to establish self-supervised constraints, enabling the model to learn the illumination and reflectance components based on the information contained within the low-light images themselves. Therefore, it can reduce the costs of acquiring, registering, and labeling paired low-light and normal-light images, making it particularly suitable for applications such as monitoring, industrial inspection, and mobile imaging where it is difficult to repeatedly capture normal-light reference images.

[0070] 4. It helps preserve the color, texture, and structural details of the image. Corresponding to L-Net, R-Net, and the reflection consistency loss in this invention, the model decomposes the image into illumination and reflection components. The illumination component primarily describes the image's brightness distribution, while the reflection component primarily describes the object's color, texture, edges, and structural information.

[0071] By constraining the reflectance components R_1 and R_2 of two low-light images of the same scene to remain consistent, the model can attribute the differences between the two images primarily to changes in illumination, rather than changes in the inherent content of objects. Therefore, while improving image brightness, it can better preserve the original scene structure, object outlines, colors, and textures, reducing structural blurring and detail loss that may occur with traditional enhancement methods.

[0072] 5. It can suppress random noise and inappropriate image features. Corresponding to the P-Net projection processing and projection loss in this invention, before performing Retinex decomposition, the low-light image is first processed by P-Net to generate an intermediate projected image. Under the constraint of projection loss, P-Net retains the main content of the original image and filters out some sensor noise, random noise, and interference features that are not suitable for direct Retinex decomposition.

[0073] Furthermore, random noise in different low-light images of the same scene is usually inconsistent. This invention, however, uses reflection consistency loss to ensure that the reflection components of two images remain consistent, making it less likely for the network to learn random noise in a single image as stable scene reflection information. Therefore, this invention can improve image brightness while reducing the possibility of noise being amplified simultaneously.

[0074] 6. It can improve problems such as uneven lighting, local overexposure, and lighting artifacts. Corresponding to the Retinex reconstruction loss and illumination smoothing constraint in this invention, this invention utilizes the product of the reflection component and the illumination component to reconstruct the intermediate projection image, so that the estimated two components satisfy the Retinex image formation relationship.

[0075] Meanwhile, by limiting the predicted illumination components from deviating from a reasonable brightness distribution through initial illumination constraints and by maintaining spatial smoothness in the illumination map through illumination gradient constraints, the possibility of texture information being incorrectly preserved in the illumination components is reduced. Independently correcting the illumination components before recombinating them with the reflection components helps to obtain more uniform and natural enhanced brightness, and reduces local overexposure, obvious halos, and illumination boundary artifacts.

[0076] 7. Enhancement methods can be flexibly selected according to actual application needs. Corresponding to the rapid adaptation and image output process for new tasks in this invention, when there are low-light samples of new scenes, a few updates can be made based on the meta-training parameters to obtain model parameters that are more suitable for the current scene; when there are few adaptation samples or when real-time requirements are high, the meta-training parameters can also be used directly for image enhancement.

[0077] Therefore, this invention allows for a flexible choice between enhanced quality and processing speed, making it suitable for both offline image processing and applications requiring rapid deployment or real-time processing, such as dynamic video surveillance, mobile terminals, vehicle vision, and industrial inspection.

[0078] 8. It has good objective image enhancement effect. Based on the comparative experimental results, the present invention achieves a PSNR of 19.57 and an LPIPS of 0.27 on the LOL dataset; and a PSNR of 21.29 and an SSIM of 0.81 on the SICE dataset. Among the listed comparative methods, the present invention achieves the highest PSNR on both datasets, the highest SSIM on the SICE dataset, and the lowest LPIPS on the LOL dataset, demonstrating that the present invention has good overall performance in image distortion control, structure preservation, and perceptual quality.

[0079] In summary, this invention, through the synergistic effect of intra-task decomposition learning in PairLIE and inter-task meta-learning in MAML, enables the low-light image enhancement model to achieve strong fast adaptation capability, cross-scene generalization capability, and enhancement stability without relying on a large number of normal illumination paired images. At the same time, it can better preserve image details, suppress noise, and improve the enhancement effect under complex lighting conditions.

[0080] Corresponding to the fast low-light image enhancement method based on PairLIE and meta-learning described in the above embodiments, this invention also provides a fast low-light image enhancement system based on PairLIE and meta-learning, comprising: The task construction module is used to construct multiple low-light image enhancement tasks from different low-light images. Each task is sampled from the low-light task distribution p(T), denoted as: {T i} i =1 ~p(T), where T i Let p(T) represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-dimensional training process. The task learning module is used to perform in-task learning within the PairLIE network for each low-light image enhancement task, obtaining the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are...i It is used for inter-task meta-learning, and the common initial parameter θ0 is optimized across tasks to obtain the meta-training parameter θ_T; The image enhancement module is used to enhance the image of a new low-light scene that was not included in the training when a new low-light scene appears after meta-training. It uses the meta-training parameter θ_T as the initial parameter of the model, and only needs to use a small number of low-light images of the new scene and perform a few parameter updates to obtain an enhancement model suitable for the new scene and perform image enhancement on the low-light images of the new scene. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure or illumination distribution.

[0081] It should be noted that the information interaction and execution process between the above modules / units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0083] This invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the fast low-light image enhancement method based on PairLIE and meta-learning provided in the first aspect.

[0084] In applications, terminal devices may include, but are not limited to, processors and memory. These are merely examples of terminal devices and do not constitute a limitation on them. They may include more or fewer components, combinations of certain components, or different components, such as input / output devices and network access devices. Input / output devices may include cameras, audio capture / playback devices, displays, etc. Network access devices may include network modules for wireless network communication with external devices.

[0085] In applications, the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0086] In applications, the memory may be an internal storage unit of the terminal device in some embodiments, such as the hard drive or RAM of the terminal device. In other embodiments, the memory may be an external storage device of the terminal device, such as a plug-in hard drive, a smart media card (SMC), or a flash card. The memory may also include both internal and external storage units of the terminal device. The memory is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of a computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0087] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps in the above-described method embodiments.

[0088] The present invention implements all or part of the processes in the methods of the above embodiments by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, such as a USB flash drive, a portable hard drive, a magnetic disk, or an optical disk.

[0089] Those skilled in the art will recognize that the device and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0090] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be through some interfaces, or indirect couplings or communication connections between devices, and may be electrical, mechanical, or other forms.

[0091] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A fast low-light image enhancement method based on PairLIE and meta-learning, characterized in that, include: Different low-light images are constructed into multiple low-light image enhancement tasks. Each task samples from the low-light task distribution p(T), denoted as: {T i } i =1 ~p(T), where T i Let p(T) represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-dimensional training process. For each low-light image enhancement task, in-task learning is performed in the PairLIE network to obtain the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are... i It is used for inter-task meta-learning, and the common initial parameter θ0 is optimized across tasks to obtain the meta-training parameter θ_T; After meta-training, when a new low-light scene that was not included in the training appears, the meta-training parameter θ_T is used as the initial parameter of the model. Only a small number of low-light images of the new scene are needed and a few parameter updates are performed to obtain an enhanced model suitable for the new scene and perform image enhancement on the low-light images of the new scene. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure or illumination distribution.

2. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 1, characterized in that, The in-task learning part of PairLIE includes a pair of low-light image input units, a projection network P-Net, an illumination estimation network L-Net, a reflection estimation network R-Net, a loss calculation unit, and an in-task parameter update unit. The parameters of P-Net, L-Net, and R-Net together constitute the parameters to be learned: θ = {θ_P, θ_L, θ_R} in, θ_P, θ_L, and θ_R represent the network parameters of P-Net, L-Net, and R-Net, respectively. The networks of the same type in the two image processing branches have the same network structure and share the corresponding parameters to ensure that the pairs of low-light images are processed according to the same projection and decomposition rules.

3. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 2, characterized in that, P-Net, L-Net, and R-Net are trained together using projection loss, reflection consistency loss, and Retinex reconstruction loss. The three types of losses are used to constrain the fidelity of projection information, the consistency of paired image reflection components, and the reconstructibility of illumination-reflection decomposition, respectively.

4. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 1, characterized in that, Within a specific low-light task, the task-specific parameter θ is learned based on the PairLIE total loss. i For the i-th low-light enhancement task, the PairLIE network is first initialized with the common initial parameter θ0: i i (0) = θ0 Then, the total in-task loss is calculated using paired low-light images from the task, and one or more gradient updates are performed: i i (s+ ¹ ) = θ i (s) a _{θ i (s) } _total(T i ;θ i (s) ) Where α represents the in-task learning rate, and s represents the number of in-task updates; after a preset number of updates K, the task-specific parameter θ corresponding to the i-th task is obtained. i =θ i (K) ; Different lighting tasks yielded θ1, θ2, ..., θ i Task-specific parameters reflect the model's parameter state after minor updates to the current lighting conditions.

5. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 1, characterized in that, The inter-task meta-learning part optimizes the common initial parameter θ0 in reverse based on the adaptation loss after a small number of updates for multiple tasks, so that the common initial parameter can converge quickly with a small number of gradient updates when facing different low-light tasks. The meta-loss is obtained by summing or averaging the query set losses of multiple tasks in a meta-training batch. _meta = (1 / N) Σ i=1 _total(Q i ;θ i ) Update the common initial parameters based on the meta-loss: θ0 ← θ0 b _{θ0} _meta Where β represents the step size of the inter-task meta-update; when using the full MAML update method, the gradient of the meta-loss is backpropagated to the common initial parameters through the intra-task parameter update process; the inter-task meta-parameter update uses the Adam optimizer, which repeatedly executes task sampling, intra-task update, query set loss calculation and inter-task meta-update until the meta-loss converges or reaches the preset number of training rounds, to obtain the final meta-training parameters θ_T.

6. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 1, characterized in that, After meta-training is complete, when the model encounters a new low-light scene T_new that did not participate in meta-training, it first loads the meta-training parameters as the initial parameters for the new task: θ_new (0) = θ_T Subsequently, a small number of paired low-light images from the new scene are acquired, and a few sub-gradient updates are performed using the same projection loss, reflection consistency loss, and Retinex reconstruction loss as in the meta-training phase: θ_new (s+ ¹ ) = θ_new (s) a_adapt _{θ_new (s) } _total(T_new) After a limited number of updates, the parameter θ_new suitable for the new low-light scene is obtained.

7. The fast low-light image enhancement method based on PairLIE and meta-learning as described in claim 2, characterized in that, For the low-light image I_low to be enhanced, it is first input into P-Net to obtain the intermediate projected image i: i = P(I_low;θ_P) Then, the intermediate projected image is input into L-Net and R-Net respectively to obtain the illumination component L and the reflection component R: L = L(i; θ_L), R = R(i; θ_R) Nonlinear correction of the illumination component: L′ = L^λ Where λ is the illumination correction factor; for illumination maps with pixel values ​​normalized to the range of [0, 1], when 0 < λ < 1, the value of the illumination component can be increased, thereby brightening low-illuminance areas. The corrected illumination component is multiplied pixel by pixel with the reflection component to obtain the enhanced image: I_enhanced = R ⊙ L′ = R ⊙ L^λ Finally, the range of pixel values ​​in the enhanced image is limited: I_out = clip(I_enhanced, 0, 1) Converting I_out to an 8-bit image, a 16-bit image, or a floating-point image yields an enhanced result that improves brightness while preserving color, texture, and structural information.

8. A fast low-light image enhancement system based on PairLIE and meta-learning, characterized in that, include: The task construction module is used to construct multiple low-light image enhancement tasks from different low-light images. Each task is sampled from the low-light task distribution p(T), denoted as: {T i } i =1 ~p(T), where T i Let p(T) represent the i-th low-light image enhancement task, p(T) represent the low-light task distribution, and N represent the number of tasks sampled during the first-dimensional training process. The task learning module is used to perform in-task learning within the PairLIE network for each low-light image enhancement task, obtaining the task-specific parameters θ corresponding to the current task. i Then, the θ1, θ2, ..., θ values ​​obtained from multiple tasks are... i It is used for inter-task meta-learning, and the common initial parameter θ0 is optimized across tasks to obtain the meta-training parameter θ_T; The image enhancement module is used to enhance the image of a new low-light scene that was not included in the training when a new low-light scene appears after meta-training. The meta-training parameter θ_T is used as the initial parameter of the model. Only a small number of low-light images of the new scene are needed and a few parameter updates are performed to obtain an enhancement model suitable for the new scene and enhance the low-light images of the new scene. Each low-light image enhancement task includes at least one pair of low-light images I1 and I2 in the same scene. The two images have the same or similar scene structure and object reflection properties, but have different brightness, exposure or illumination distribution.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the fast low-light image enhancement method based on PairLIE and meta-learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the fast low-light image enhancement method based on PairLIE and meta-learning as described in any one of claims 1 to 7.

Images