Low-light image enhancement method and device

Abstract

The invention is suitable for the field of computer vision, and provides a low-light image enhancement method and device. In the embodiment of the invention, Poisson noise perception Retinex decomposition is firstly carried out on a target low-light image needing to be enhanced, components generated by decomposition are input into a neural network trained based on a decomposition result so as to obtain a feature map of the low-light image, and finally, an illumination feature map and a reflectivity feature map in the feature map are multiplied so as to obtain an enhanced low-light image. And noise suppression is carried out on the multiplication result through a Poisson noise feature map, so that an enhanced image of the target low-light image is obtained. According to the method, a traditional Retinex model is expanded into a three-component decomposition mode containing a Poisson noise component through a Poisson noise perception Retinex decomposition framework, and the decomposition result is processed through the neural network integrating the reconstruction loss, the Retinex estimation loss and the noise estimation loss, so that the quality of the enhanced low-light image can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of computer vision, and more particularly to a method and apparatus for low-light image enhancement. Background Technology

[0002] Image enhancement and denoising techniques for extreme low-light environments can be widely applied in scenarios with high requirements for image visual quality and information integrity, such as photography, video surveillance, autonomous driving, and medical imaging. In low-light or extremely low-light environments, image acquisition devices often suffer from severe quality degradation due to insufficient photon counts. This degradation manifests as high noise interference, low contrast, and distorted illumination distribution. Such degraded images not only severely affect the visual perception of the human eye but also lead to a significant decrease in the performance of downstream tasks that rely on image data (such as object detection, image segmentation, and feature recognition). For example, in autonomous driving scenarios, low-light images may prevent vehicles from accurately identifying pedestrians or traffic signs, and in medical imaging scenarios, crucial lesion details may be missed. Therefore, low-light image enhancement and denoising has become a critical technical problem that urgently needs to be solved in the field of computer vision.

[0003] Existing low-light image enhancement and denoising techniques generally suffer from defects such as insufficient Poisson noise modeling, poor synergy between enhancement and denoising, difficulty in ensuring color consistency, and low computational efficiency, resulting in quality degradation of the enhanced low-light image. Summary of the Invention

[0004] In view of this, this application provides a low-light image enhancement method and apparatus to solve the problems of insufficient Poisson noise modeling, poor synergy between enhancement and denoising, and difficulty in ensuring color consistency, thereby improving the quality of the enhanced low-light image.

[0005] A first aspect of this application provides a low-light image enhancement method, the method comprising: The target low-light image that needs to be enhanced is subjected to Poisson noise-aware Retinex decomposition to generate the decomposition result of the target low-light image, wherein the decomposition result includes illuminance component, reflectance component and Poisson noise component. The neural network is trained using the decomposition results, and the feature map of the target low-light image is obtained through the trained neural network. The feature map includes an illuminance feature map, a reflectance feature map, and a Poisson noise feature map. The illuminance feature map and reflectance feature map are multiplied to obtain a preliminary enhanced image. Then, the noise of the preliminary enhanced image is suppressed by the Poisson noise feature map to obtain the target enhanced image of the target low-light image.

[0006] Optionally, performing Poisson noise-aware Retinex decomposition on the low-light image of the target to be enhanced includes: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

[0007] Optionally, training the neural network using the decomposition results includes: The reconstruction loss term of the neural network is obtained through the decomposition results. Defined as Where A is the low-light image of the target, C is the illuminance component, D is the reflectance component, and E is the Poisson noise component; The noise estimation loss term Defined as , where δ is a small constant that ensures computational stability; Retinex estimate the loss term Defined as Where η is the balance coefficient. For component loss terms, To smooth out the loss term; Define the noise estimation loss term as Where δ is a small constant that ensures computational stability; The composite loss function of the neural network is determined as follows: , where k1 and k2 are weighting coefficients.

[0008] Optionally, the neural network includes three parallel branches, and obtaining the feature map of the target low-light image through the trained neural network includes: The first branch uses the sigmoid activation function to generate a single-channel illuminance feature map; The second branch uses the sigmoid activation function to generate a three-channel reflectance feature map; The third branch uses the tanh activation function to generate a three-channel Poisson noise feature map.

[0009] Optionally, the neural network is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

[0010] A second aspect of this application provides a low-light image enhancement apparatus, the apparatus comprising: The image decomposition unit is used to perform Poisson noise-aware Retinex decomposition on the target low-light image that needs to be enhanced, and generate the decomposition result of the target low-light image, wherein the decomposition result includes an illuminance component, a reflectance component and a Poisson noise component. The feature map extraction unit is used to train a neural network through the decomposition results and obtain the feature map of the target low-light image through the trained neural network, wherein the feature map includes an illuminance feature map, a reflectance feature map and a Poisson noise feature map; The image enhancement unit is used to multiply the illuminance feature map and the reflectance feature map to obtain a preliminary enhanced image, and then use the Poisson noise feature map to suppress noise in the preliminary enhanced image to obtain the target enhanced image of the target low-light image.

[0011] Optionally, the image decomposition unit performing Poisson noise-aware Retinex decomposition on the target low-light image requiring enhancement includes: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

[0012] Optionally, training the neural network using the decomposition results in the feature map extraction unit includes: The reconstruction loss term of the neural network is obtained through the decomposition results. Defined as Where A is the low-light image of the target, C is the illuminance component, D is the reflectance component, and E is the Poisson noise component; The noise estimation loss term Defined as , where δ is a small constant that ensures computational stability; Retinex estimate the loss term Defined as Where η is the balance coefficient. For component loss terms, To smooth out the loss term; Define the noise estimation loss term as Where δ is a small constant that ensures computational stability; The composite loss function of the neural network is determined as follows: , where k1 and k2 are weighting coefficients.

[0013] Optionally, the neural network in the feature map extraction unit contains three parallel branches; The first branch uses the sigmoid activation function to generate a single-channel illuminance feature map; The second branch uses the sigmoid activation function to generate a three-channel reflectance feature map; The third branch uses the tanh activation function to generate a three-channel Poisson noise feature map.

[0014] Optionally, the neural network in the feature map extraction unit is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

[0015] In the embodiments provided in this application, for a target low-light image that needs enhancement, a Poisson noise-aware Retinex decomposition is first performed. The decomposed components are then input into a neural network trained based on the decomposition results to obtain a feature map of the low-light image. Finally, the illuminance feature map and reflectance feature map in the feature map are multiplied, and noise suppression is applied to the multiplication result using the Poisson noise feature map to obtain the enhanced image of the target low-light image. This application extends the traditional Retinex model to a three-component decomposition including Poisson noise components through a Poisson noise-aware Retinex decomposition framework. By processing the decomposition results through a neural network integrating reconstruction loss, Retinex estimation loss, and noise estimation loss, the quality of the enhanced low-light image can be effectively improved. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the method provided in this application embodiment; Figure 2 A schematic diagram of the loss function design provided for embodiments of this application; Figure 3 An enhanced method architecture diagram provided for embodiments of this application; Figure 4 Comparison images of visual enhancement effects provided for embodiments of this application; Figure 5 This is a structural diagram of the device provided in the embodiments of this application; Figure 6 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0020] This application first describes existing low-light image enhancement methods.

[0021] Existing technologies are mainly divided into two categories: classical image processing methods and deep learning methods. However, both types of methods have significant drawbacks when dealing with extreme low-light scenes.

[0022] 1. Classic Low-Light Image Enhancement and Denoising Methods. This method is based on histogram adjustment, which improves contrast and brightness by redistributing the gray levels of image pixels. Typical techniques include histogram equalization and its improved versions. However, these methods do not consider the noise characteristics of low-light images. While enhancing brightness, they tend to amplify noise. In extreme low-light environments, they can produce severe over-enhancement artifacts and cannot effectively restore the true structure and color information of the image. Retinex-based methods decompose images into reflectance and illuminance components, and enhance them by adjusting the illuminance component. Representative techniques include single-scale Retinex, multi-scale Retinex, and LIME. LIME achieves pixel-level brightness amplification by estimating a structure-aware illuminance map, but it does not incorporate denoising into its core algorithm framework and relies solely on an external denoiser for post-processing. This results in poor adaptability to different noise intensities and image structures. In extreme low-light scenes, the "enhance first, then denoise" process further exacerbates noise pollution. Moreover, traditional Retinex methods generally assume that the noise is zero-mean additive Gaussian noise, which does not match the signal-dependent Poisson noise characteristics that dominate in actual low-light scenes, leading to low decomposition accuracy and limited enhancement effects.

[0023] 2. Deep Learning-Based Low-Light Image Enhancement and Denoising Methods. While this method is gradually becoming mainstream, key issues remain. Some methods, such as RetinexNet, employ an "enhancement first, then denoising" or "independent optimization of enhancement and denoising" approach, which fails to adapt to the strong correlation between Poisson noise and signal strength. Weak signals in low-light regions are easily overwhelmed by noise, leading to image structure distortion or color shift. Methods like Zero-DCE and EnlightenGAN do not explicitly model the noise distribution in low-light environments, ignoring the dominant role of Poisson noise in extreme low-light scenes, resulting in overexposure and color distortion in the enhanced image. While methods based on visual Transformers can improve structure preservation, their large number of model parameters and high computational complexity make them difficult to meet the low-latency requirements of real-time applications.

[0024] This application provides a low-light image enhancement method and apparatus to solve the problems of insufficient Poisson noise modeling, poor synergy between enhancement and denoising, and difficulty in ensuring color consistency in the prior art, thereby improving the quality of the enhanced low-light image.

[0025] The technical solutions of this application will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0026] like Figure 1 The diagram shown is a flowchart of a low-light image enhancement method provided in this application. The process may include the following steps: Step S101: Perform Poisson noise-aware Retinex decomposition on the target low-light image that needs to be enhanced to generate the decomposition result of the target low-light image.

[0027] In this embodiment, the decomposition result includes an illuminance component, a reflectance component, and a Poisson noise component. Based on the physical characteristics of Poisson noise, the target low-light image A to be enhanced needs to be represented as an element-wise product of illuminance component C, reflectance component D, and Poisson noise component E. The illuminance component C can be estimated first using low-pass filtering or illumination estimation techniques, and then the reflectance component D can be separated from A using Retinex theory or logarithmic domain decomposition. For the Poisson noise component E, it is first assumed that the noise of each pixel follows a Poisson distribution, and then... Determine the ideal signal B, where, This is an element-wise multiplication. Then, a Poisson random number generator is used to generate a Poisson noise component E based on the ideal signal B.

[0028] In another embodiment, performing Poisson noise-aware Retinex decomposition on the low-light image of the target to be enhanced includes: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

[0029] In this embodiment, the true pixel value at each location can be estimated using a filtering method. Then, combined with the observed count value of the entire low-light image of the target (i.e., the total number of pixels), the Poisson noise component value E(x) of each pixel is determined, thereby determining the Poisson noise component E. Then, based on a mathematical model... The process of determining the specific illuminance component C and reflectance component D for Poisson noise perception Retinex decomposition is the same as in the previous embodiment, and will not be repeated here.

[0030] Step S102: Train a neural network using the decomposition results, and obtain the feature map of the target low-light image using the trained neural network.

[0031] In this embodiment, the neural network needs to be trained first using the decomposition results of the low-light image. The training process is as follows: Figure 2 As shown: 1. Reconstruct the loss term using the target low-light image A, illuminance component C, reflectance component D, and Poisson noise component E. This is used to ensure the accuracy of the decomposition.

[0032] 2. During training, the target low-light image A is initially enhanced to generate an initial enhanced image B. Then, the illuminance component C and reflectance component D are combined to determine the Retinex estimation loss term. This is used to ensure the consistency of reflectivity in the enhanced image. In the above formula, η is a manually set balancing parameter, which can be set to 0.9. For component loss terms, use the formula Sure. To smooth out the remaining loss term, by ∇C and ∇D represent the gradients of illuminance and reflectance in the horizontal and vertical directions, respectively, and θ is a parameter for balancing structure and smoothness, which can be manually set to 10.

[0033] 3. Determine the noise estimation loss term using the aforementioned low-light image A of the target. This formula is specifically designed to suppress Poisson noise. The δ value in this formula is a small constant of 10^-6, ensuring computational stability.

[0034] 4. Based on the reconstruction loss term, Retinex estimation loss term, and noise estimation loss term, the composite loss function of the above neural network is determined as follows: k1 and k2 are weight coefficients, which can be manually set to 0.5 and 0.1, respectively. The composite loss function is then used for iterative optimization training to obtain the neural network trained based on the decomposition results.

[0035] Since the composite loss function integrates reconstruction loss, Retinex estimation loss and noise estimation loss, it can effectively solve the color distortion problem under extreme low light conditions by using physics-driven Poisson noise modeling, and significantly improve visual quality and color consistency while maintaining image structural details.

[0036] After training, the neural network is used to extract feature maps, which include illuminance feature maps, reflectance feature maps, and Poisson noise feature maps. The specific extraction process is as follows: Figure 3 As shown: The initial feature extraction layer of this neural network uses a 3×3 convolutional kernel with 64 output channels to extract basic features from the input RGB image. The encoding path contains three consecutive convolutional layers with a stride of 2, each followed by a ReLU activation function to extract multi-scale hierarchical features. The decoding path contains three transposed convolutional layers, which concatenate the features of corresponding layers from the encoder through skip connections to achieve spatial resolution restoration and detail preservation.

[0037] The output of this neural network consists of three parallel branches: the first branch uses the sigmoid activation function to generate a single-channel illuminance feature map, the second branch uses the sigmoid activation function to generate a three-channel reflectance feature map, and the third branch uses the tanh activation function to generate a three-channel noise feature map.

[0038] This embodiment uses a parallel branch structure to process illumination, reflectivity, and noise components simultaneously, achieving coordinated optimization of image enhancement and denoising, and avoiding the noise amplification problem in the traditional "enhancement first, then denoising" process.

[0039] In another embodiment, the neural network is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

[0040] This embodiment can use a publicly available low-light (LOL) image dataset for training and testing, instead of the decomposition results of the target low-light image described above. The dataset contains 500 pairs of low-light and corresponding normal-light images, acquired by varying exposure time and ISO while keeping other camera settings constant. During training, the decomposition results of the target low-light image are replaced with the low-light decomposition results from this dataset. When training the neural network, the formula is used to determine the component loss term in the Retinex estimation loss term. The initial enhanced image B in the image is replaced with a normal illumination image corresponding to low illumination.

[0041] Step S103: Multiply the illuminance feature map and the reflectance feature map to obtain a preliminary enhanced image, and then use the Poisson noise feature map to suppress noise in the preliminary enhanced image to obtain the target enhanced image of the target low-light image.

[0042] In this embodiment, element-wise multiplication can be used to multiply the illuminance feature map and the reflectance feature map. This couples the global illumination intensity information provided by the illuminance map with the detailed texture information carried by the reflectance map at the pixel level, thereby synthesizing a preliminary enhanced image.

[0043] Then, the Poisson noise feature map E is used to suppress noise in the initially enhanced image. This process requires incorporating the statistical characteristics of the noise map into the enhancement process. First, E is considered as a weight map reflecting the confidence level of noise intensity at each pixel. Typically, the intensity of Poisson noise is related to the signal intensity. Therefore, E can be inverted or transformed to generate a noise suppression weight matrix W. Then, an adaptive filtering process based on W is used to optimize the initially enhanced image. In this way, in areas with significant noise (large E value, small W value), smoothing is relied upon more to suppress noise, while in areas with high signal-to-noise ratio (small E value, large W value), more enhancement details are preserved, achieving content-aware adaptive noise reduction.

[0044] After noise suppression through the above steps, the enhanced image of the target low-light image can be obtained.

[0045] This concludes the process. Figure 1 The process is shown below.

[0046] In this embodiment, for a target low-light image that needs to be enhanced, a Poisson noise-aware Retinex decomposition is first performed on it, and the components generated by the decomposition are input into a neural network trained based on the decomposition result to obtain a feature map of the low-light image. Finally, the illuminance feature map and reflectance feature map in the feature map are multiplied, and the multiplication result is noise-suppressed by the Poisson noise feature map to obtain an enhanced image of the target low-light image.

[0047] The beneficial effects of this application are at least as follows: (1) This invention proposes a Poisson noise-aware Retinex decomposition framework, which extends the traditional Retinex model into a three-component decomposition that includes Poisson noise components, better conforms to the physical characteristics of extreme low-light scenes, and provides a solid theoretical basis for image enhancement and denoising.

[0048] (2) The present invention designs an end-to-end convolutional encoder-decoder architecture, which processes illumination, reflectivity and noise components simultaneously through a parallel branch structure, thereby achieving coordinated optimization of image enhancement and denoising, and avoiding the problem of noise amplification in the traditional "enhancement first and then denoising" process.

[0049] (3) The composite loss function proposed in this invention integrates reconstruction loss, Retinex estimation loss and noise estimation loss. Through physical-driven Poisson noise modeling, it effectively solves the color distortion problem under extreme low light conditions and significantly improves visual quality and color consistency while maintaining image structural details.

[0050] To further illustrate the effects of the present invention, a specific embodiment is provided, which compares multiple methods under the same experimental conditions. The experimental results demonstrate the superior performance of the present invention in terms of visual quality and quantitative indicators.

[0051] In terms of visual quality, such as Figure 4 As shown, this invention can generate enhanced images with consistent colors and rich details under different lighting levels (the lighting gradually decreases from the first row to the fourth row). Compared with existing methods such as LIME, Deep-Retinex, and Zero-DCE, this invention can maintain good color consistency even under extremely low lighting conditions, while other methods show obvious green shifts or other color distortions.

[0052] To further verify color consistency, we analyzed the histogram distribution of different methods across the RGB channels. An image with good color consistency should exhibit a similar histogram distribution across the three RGB channels. It can be seen that the image generated by the method of this invention has a highly similar histogram across the RGB channels, while other methods show obvious channel imbalance, confirming the advantage of this invention in maintaining color consistency.

[0053] For quantitative metrics, we used three image quality evaluation indicators: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and No-Reference Image Quality Assessment (NIQE). Higher PSNR and SSIM indicate better image quality, while lower NIQE indicates better image quality. The experimental results are shown in the table below: method PSNR (dB) SSIM NIQE LIME 15.66 0.42 9.35 Deep Retinex 16.06 0.58 4.31 Zero-DCE 12.35 0.55 3.42 PNAR 16.53 0.583 - IQA 17.95 0.658 - This invention 17.12 0.67 3.54 The results in the table show that this invention achieved the highest score on the SSIM metric, indicating a significant advantage in structure preservation; it ranked second only to the IQA method on the PSNR metric, demonstrating excellent performance in pixel-level accuracy; and it ranked second only to Zero-DCE on the NIQE metric, indicating that the generated images are very close to natural images in terms of visual perception quality. In summary, this invention achieves high-quality image enhancement while maintaining color consistency, especially excelling under extremely low-light conditions.

[0054] This application also provides a low-light image enhancement device, such as... Figure 5 As shown, the device includes: Image decomposition unit 501 is used to perform Poisson noise-aware Retinex decomposition on the target low-light image that needs to be enhanced, and generate the decomposition result of the target low-light image, wherein the decomposition result includes illuminance component, reflectance component and Poisson noise component. The feature map extraction unit 502 is used to train a neural network through the decomposition result and obtain the feature map of the target low-light image through the trained neural network, wherein the feature map includes an illuminance feature map, a reflectance feature map and a Poisson noise feature map; The image enhancement unit 503 is used to multiply the illuminance feature map and the reflectance feature map to obtain a preliminary enhanced image, and then use the Poisson noise feature map to suppress noise in the preliminary enhanced image to obtain a target enhanced image of the target low-light image.

[0055] In another embodiment, the Poisson noise-aware Retinex decomposition of the target low-light image to be enhanced in the image decomposition unit includes: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

[0056] In another embodiment, training the neural network using the decomposition results in the feature map extraction unit includes: The reconstruction loss term of the neural network is obtained through the decomposition results. Defined as Where A is the low-light image of the target, C is the illuminance component, D is the reflectance component, and E is the Poisson noise component; The noise estimation loss term Defined as , where δ is a small constant that ensures computational stability; Retinex estimate the loss term Defined as Where η is the balance coefficient. For component loss terms, To smooth out the loss term; Define the noise estimation loss term as Where δ is a small constant that ensures computational stability; The composite loss function of the neural network is determined as follows: , where k1 and k2 are weighting coefficients.

[0057] In another embodiment, the neural network in the feature map extraction unit comprises three parallel branches; The first branch uses the sigmoid activation function to generate a single-channel illuminance feature map; The second branch uses the sigmoid activation function to generate a three-channel reflectance feature map; The third branch uses the tanh activation function to generate a three-channel Poisson noise feature map.

[0058] In another embodiment, the neural network in the feature map extraction unit is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

[0059] The above embodiments of the present invention provide a low-light image enhancement method and a low-light image enhancement device based on the method. The quality of the enhanced low-light image can be improved by the above method and device.

[0060] This embodiment also discloses a computer device, such as... Figure 6 As shown, the computer device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement any of the low-light image enhancement methods described above.

[0061] Furthermore, in the above-described embodiments of the low-light image enhancement device, the logical division of each program module is merely illustrative. In practical applications, the functions described above can be assigned to different program modules as needed, for example, for the sake of hardware configuration requirements or software implementation convenience. That is, the internal structure of the low-light image enhancement device can be divided into different program modules to complete all or part of the functions described above.

[0062] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A low-light image enhancement method, characterized in that, The method includes: The target low-light image that needs to be enhanced is subjected to Poisson noise-aware Retinex decomposition to generate the decomposition result of the target low-light image, wherein the decomposition result includes illuminance component, reflectance component and Poisson noise component. The neural network is trained using the decomposition results, and the feature map of the target low-light image is obtained through the trained neural network. The feature map includes an illuminance feature map, a reflectance feature map, and a Poisson noise feature map. The illuminance feature map and reflectance feature map are multiplied to obtain a preliminary enhanced image. Then, the noise of the preliminary enhanced image is suppressed by the Poisson noise feature map to obtain the target enhanced image of the target low-light image.

2. The method according to claim 1, characterized in that, The step of performing Poisson noise-aware Retinex decomposition on the target low-light image that needs enhancement includes: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

3. The method according to claim 1, characterized in that, The step of training the neural network using the decomposition results includes: The reconstruction loss term of the neural network is obtained through the decomposition results. Defined as Where A is the low-light image of the target, C is the illuminance component, D is the reflectance component, and E is the Poisson noise component; The noise estimation loss term Defined as , where δ is a small constant that ensures computational stability; Retinex estimate the loss term Defined as Where η is the balance coefficient. For component loss terms, To smooth out the loss term; Define the noise estimation loss term as Where δ is a small constant that ensures computational stability; The composite loss function of the neural network is determined as follows: , where k1 and k2 are weighting coefficients.

4. The method according to claim 1, characterized in that, The neural network contains three parallel branches, and the process of obtaining the feature map of the target low-light image through the trained neural network includes: The first branch uses the sigmoid activation function to generate a single-channel illuminance feature map; The second branch uses the sigmoid activation function to generate a three-channel reflectance feature map; The third branch uses the tanh activation function to generate a three-channel Poisson noise feature map.

5. The method according to claim 1, characterized in that, The neural network is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

6. A low-light image enhancement device, characterized in that, The device includes: The image decomposition unit is used to perform Poisson noise-aware Retinex decomposition on the target low-light image that needs to be enhanced, and generate the decomposition result of the target low-light image, wherein the decomposition result includes an illuminance component, a reflectance component and a Poisson noise component. The feature map extraction unit is used to train a neural network through the decomposition results and obtain the feature map of the target low-light image through the trained neural network, wherein the feature map includes an illuminance feature map, a reflectance feature map and a Poisson noise feature map; The image enhancement unit is used to multiply the illuminance feature map and the reflectance feature map to obtain a preliminary enhanced image, and then use the Poisson noise feature map to suppress noise in the preliminary enhanced image to obtain the target enhanced image of the target low-light image.

7. The apparatus according to claim 6, characterized in that, The image decomposition unit performs Poisson noise-aware Retinex decomposition on the target low-light image that needs enhancement, including: Through formula Determine the Poisson noise component E, where, Let E be the value of the Poisson noise component at pixel j, B(j) be the true pixel value of the low-light image of the target at pixel j, and m represent the observation count value. Based on mathematical model A Poisson noise-sensing Retinex decomposition is performed, where A is the low-light image of the target, C is the illuminance component, and D is the reflectance component.

8. The apparatus according to claim 6, characterized in that, The feature map extraction unit includes training a neural network using the decomposition results, which includes: The reconstruction loss term of the neural network is obtained through the decomposition results. Defined as Where A is the low-light image of the target, C is the illuminance component, D is the reflectance component, and E is the Poisson noise component; The noise estimation loss term Defined as , where δ is a small constant that ensures computational stability; Retinex estimate the loss term Defined as Where η is the balance coefficient. For component loss terms, To smooth out the loss term; Define the noise estimation loss term as Where δ is a small constant that ensures computational stability; The composite loss function of the neural network is determined as follows: , where k1 and k2 are weighting coefficients.

9. The apparatus according to claim 6, characterized in that, The neural network in the feature map extraction unit contains three parallel branches; The first branch uses the sigmoid activation function to generate a single-channel illuminance feature map; The second branch uses the sigmoid activation function to generate a three-channel reflectance feature map; The third branch uses the tanh activation function to generate a three-channel Poisson noise feature map.

10. The apparatus according to claim 6, characterized in that, The neural network in the feature map extraction unit is trained and tested using a preset dataset, which includes several low-light images and normal-light images corresponding to the low-light images.

Images