A multi-exposure based fusion generated low-light image enhancement system

Abstract

The application discloses a kind of based on multiple exposure generation fusion's low-light image enhancement system, belong to low-light image enhancement technical field.System includes: adaptive multiple exposure adjustment module (AMEA), joint prediction gamma coefficient, linear modulation factor and compensation parameter, to input low-light image carries out nonlinear and linear combination transformation, generates multiple luminance, contrast and different color response multiple exposure candidate image;Multiple exposure fusion module (MEF), original low-light image and candidate image channel splicing are carried out, and preliminary enhancement image is obtained by fine-grained weighting fusion of channel fusion coefficient and introduction of perception compensation parameter correction;Color refinement module (CRM), using encoder decoder structure and introducing fast fourier representation in bottleneck layer, realize spatial domain and frequency domain joint modeling, carry out global color calibration and local detail optimization to preliminary enhancement image.The overall network uses lightweight design such as depth separable convolution and dense connection, inference is simple, and can be run in real time in edge device.The application generates by adaptive multiple exposure, structure constraint fusion and frequency domain refinement, effectively restores dark texture, suppresses over-enhancement and color deviation, improves visual naturalness and robustness in complex low-light scene.

Description

Technical Field

[0001] This invention belongs to the field of low-light image enhancement technology, specifically relating to a low-light image enhancement system based on multi-exposure generation and fusion. Background Technology

[0002] Low-light image enhancement, as an important preprocessing step in computer vision systems, is widely used in visual tasks such as intelligent monitoring, autonomous driving, and face recognition. Its main goal is to improve the visibility, contrast, texture details, and color consistency of images in low-light environments, thereby improving the recognition accuracy and stability of downstream algorithms such as object detection, semantic segmentation, and visual tracking.

[0003] In actual imaging processes, low-light images often suffer from insufficient contrast, lack of texture in dark areas, significant noise, color distortion, and uneven brightness distribution due to factors such as underexposure, uneven ambient lighting, or limited dynamic range. To address these issues, existing low-light image enhancement methods can be broadly categorized into two types: traditional image processing methods and deep learning-based methods.

[0004] Traditional low-light image enhancement methods: Early low-light image enhancement techniques mainly included histogram equalization, gamma correction, and enhancement methods based on Retinex theory. Histogram equalization improves overall brightness and contrast by redistributing the image's gray-level histogram, but it can easily cause problems such as excessive stretching of local details, noise amplification, and visual artifacts in complex scenes. Gamma transform adjusts the image brightness distribution through nonlinear mapping, which can improve the visibility of dark areas to some extent, but its parameters are usually fixed or manually set, making it difficult to adapt to image differences under different lighting conditions. Methods based on Retinex theory enhance the image by decomposing it into illumination and reflection components, which improves the interpretability of the model to some extent, but these methods often ignore the nonlinear response of the imaging system and the noise propagation mechanism, and are prone to structural distortion and color shift in complex low-light or high dynamic range scenes. Overall, traditional methods have low computational complexity and are simple to implement, but their generalization ability and enhancement stability in complex lighting environments are limited.

[0005] Deep Learning-Based Low-Light Image Enhancement Methods: With the development of convolutional neural networks, generative adversarial networks, and Transformer architectures, end-to-end deep learning models are widely used in low-light image enhancement tasks. These methods learn the mapping relationship between low-light images and normally exposed images, achieving significant results in brightness restoration, detail enhancement, and noise suppression. Some methods incorporate Retinex theory to construct illumination-reflection separation networks to enhance the model's structural modeling capabilities. However, existing deep learning-based enhancement methods still suffer from the following problems: 1. Most methods employ a single enhancement mapping or single exposure estimation strategy. In high dynamic range or complex lighting scenarios, it is difficult to simultaneously achieve detail restoration in dark areas and preservation in bright areas, easily leading to over-enhancement, color distortion, or structural artifacts. 2. Some high-performance enhancement networks have complex structures with a large number of parameters and computational costs, making them difficult to deploy on resource-constrained edge devices, thus limiting their widespread application in practice. 3. Some methods attempt to enhance images using multi-exposure strategies, but these typically lack adaptive exposure generation mechanisms and rely heavily on fixed parameters or global fusion weights, making it difficult to model local exposure differences in fine-grained terms across spatial and channel dimensions. Furthermore, the fusion process rarely considers human visual perception preferences, potentially leading to issues such as noise, color cast, or unnatural local effects in the enhancement results.

[0006] For example, the Zero-DCE method achieves zero-reference augmentation by performing high-order nonlinear mapping on pixel values, resulting in a relatively lightweight model structure. However, it relies solely on pixel-level mapping and lacks modeling of global brightness and local structural information, leading to insufficient stability of the augmentation effect in high dynamic range scenarios. While some methods based on Transformer or complex Retinex modeling improve augmentation performance, their large model parameter scale and high computational cost make them unsuitable for real-time or embedded deployments.

[0007] In summary, existing low-light image enhancement techniques have not yet achieved an ideal balance between high dynamic range adjustment capabilities, preservation of detail and color consistency, and lightweight model deployment.

[0008] Based on the above background analysis, existing low-light image enhancement techniques suffer from the following main shortcomings in practical applications: 1. Insufficient adaptability of single enhancement strategies: Most existing methods rely on a single illumination estimation model or a single mapping function to enhance low-light images. In scenarios with complex illumination distribution, high dynamic range, or local extremely dark areas, it is difficult to simultaneously achieve dark texture restoration, highlight detail preservation, and global color consistency. Due to the lack of a multi-dimensional exposure expression mechanism, problems such as over-enhancement, loss of detail, local saturation, or color distortion easily occur during actual enhancement, resulting in insufficient stability of the enhancement results. 2. High model complexity and insufficient lightweighting: Some high-performance deep learning models typically rely on complex network structures or a large number of parameters, leading to large model size and high computational overhead, which is not conducive to deployment on edge devices with limited computing power. Under the constraint of model complexity, existing methods cannot simultaneously achieve exposure correction, texture detail restoration, high dynamic range modeling, and color consistency optimization, making it difficult to meet the needs of real-time application scenarios such as intelligent monitoring and autonomous driving. 3. Multi-exposure strategies lack adaptive generation and fine-grained fusion mechanisms: Some existing methods attempt to employ multi-exposure enhancement strategies, but they typically lack adaptive multi-exposure generation mechanisms, relying mostly on fixed exposure parameters or global weights for fusion. This fails to provide fine-grained modeling of local exposure differences in both spatial and channel dimensions. Furthermore, the fusion process rarely incorporates compensation mechanisms based on human visual perception characteristics, potentially leading to noise, artifacts, or color casts in the enhancement results. Visual naturalness and perceptual consistency need improvement. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide a low-light image enhancement system based on multi-exposure generation and fusion. Under the premise of ensuring lightweight model structure, reducing computational complexity and parameter scale, and meeting the needs of edge device deployment and real-time inference, it constructs an adaptive multi-exposure generation mechanism and a fine-grained fusion mechanism to achieve accurate modeling of the illumination distribution and dynamic range in complex low-light scenes. During the enhancement process, it effectively restores the texture details in dark areas, suppresses over-enhancement and structural distortion, and maintains the details in bright areas and the consistency of global color, thereby improving the visual naturalness and robustness of the enhancement results.

[0010] Therefore, this paper proposes a low-light image enhancement method that can achieve adaptive multi-exposure generation and fine-grained fusion while ensuring model lightweightness, and taking into account structural consistency and visual naturalness. This method has important practical application value.

[0011] A low-light image enhancement system based on multi-exposure generation fusion includes:

[0012] An adaptive multi-exposure adjustment module is used to jointly predict the gamma coefficient, linear modulation factor and compensation parameters based on the input low-light image, and perform nonlinear transformation on the low-light image to generate multiple multi-exposure candidate images with different brightness, contrast and color response. The multi-exposure fusion module is used to perform channel-by-channel stitching of the low-light image and the multi-exposure candidate image, and calculate the fusion feature image; calculate the channel-by-channel fusion coefficient based on the fusion feature image, and calculate the perception compensation parameter based on the stitched image of the low-light image and the fusion feature image; The color refinement module performs fine-grained weighted fusion of the multi-exposure candidate images using the channel-by-channel fusion coefficients, and corrects the fusion result using perceptual compensation parameters to obtain the final enhanced image.

[0013] Preferably, the adaptive multi-exposure adjustment module includes: The feature extraction unit is used to extract shallow features from the low-light image and extract deep features through a densely connected cascaded convolutional structure. Spatial attention unit (SAU) is used to predict the linear modulation factor based on the deep features. ; Global attention unit (GAU) is used to predict gamma coefficients based on the deep features. Based on linear modulation factor Predicted compensation parameters ; The adaptive multi-exposure adjustment module generates the multi-exposure candidate image according to the following formula:

[0014] in, For low-light images, This represents the candidate image under the i-th set of parameters, that is, the candidate image corresponding to different brightness, contrast and color response.

[0015] Preferably, the feature extraction unit adopts a depthwise separable convolution and dense connection structure: the first three convolutions are depthwise convolutions, the last convolution is a standard convolution, and the input of each convolution is the channel concatenation of the outputs of all previous convolutions and the shallow features.

[0016] Preferably, the global attention unit obtains global information through adaptive average pooling and performs feature mapping through two linear layers in sequence, with the channel dimension first increasing and then decreasing to the output dimension; the spatial attention unit increases the number of channels through 1×1 convolution, processes it through depthwise convolution, and then decreases the number of channels to the output dimension through 1×1 convolution.

[0017] Preferably, the process by which the multi-exposure fusion module obtains the fused feature image includes: the feature extraction unit extracts deep features from the stitched image, and then convolves them into the fused feature image using a 3×3 convolution.

[0018] Preferably, a global attention unit (GAU) is used to predict the channel-by-channel fusion coefficients. , , N To estimate the number of parameter groups.

[0019] Preferably, the perceptual compensation parameters are calculated using a global attention unit (GAU). .

[0020] Ideally, the color refinement module obtains the following results for the R, G, and B channels of the final enhanced image: (10) (11) (12) in, express The value of the m-th channel. .

[0021] Ideally, the Spatial Attention Unit (SAU) first uses 1×1 convolutions to increase the number of feature channels from C to 4C, then processes single-channel local features through depthwise convolutions, and finally reduces the number of feature channels to 3N using 1×1 convolutions. N To estimate the number of parameter groups.

[0022] Preferably, the Global Attention Unit (GAU) first acquires global information through adaptive average pooling, and then maps deep features sequentially through two linear layers; in this process, the data channel dimension first increases from C to 4C, and then decreases to 3N. N To estimate the number of parameter groups.

[0023] The present invention has the following beneficial effects: 1. Adaptive Multi-Exposure Generation Mechanism: AMEA is designed to perform a nonlinear and linear combination transformation on the input low-light image by jointly predicting the gamma coefficient, linear modulation factor, and compensation parameters. This generates multiple multi-exposure candidate images with different brightness and color responses, rather than simply performing uniform brightness scaling. This mechanism can adaptively generate multi-exposure candidate images based on the illumination and color distribution of the input image, enabling the model to simultaneously cover the needs for shadow enhancement and highlight detail preservation. This fundamentally improves the ability to depict complex, non-uniform lighting scenes, achieving synergistic optimization of shadow texture restoration, highlight suppression, and global color consistency. It solves the problems of poor adaptability and unstable enhancement effects associated with single strategies.

[0024] 2. Multi-parameter joint prediction structure: In the multi-exposure generation process, a parameter prediction structure combining global attention mechanism and spatial attention mechanism is constructed to achieve joint modeling of gamma coefficient, linear modulation factor and compensation parameter, so as to enhance the ability to express complex illumination distribution.

[0025] 3. Introducing a structurally constrained multi-exposure fusion mechanism: MEF is proposed, which constructs fusion features by channel stitching of the original low-light image and multi-exposure candidate images. The original image is used as a structural constraint to participate in the fusion process, ensuring that the fusion process is always based on real structural information and avoiding structural drift or edge blurring caused by brightness and color adjustments.

[0026] 4. Channel-level fine-grained weighting and perceptual compensation mechanism: Fine-grained weighted fusion of multi-exposure candidate images is achieved through channel-by-channel fusion coefficients. This allows the fusion process to adaptively adjust for different regions and feature responses, more accurately modeling local exposure differences at the spatial and semantic levels. Furthermore, perceptual compensation parameters are introduced to correct the fusion results, making the enhanced results more consistent with human visual perception preferences in terms of brightness distribution and color rendering. This fusion mechanism of "structural constraints + channel-level adaptive weighting + perceptual compensation" improves the efficiency of utilizing complementary information in multi-exposure candidate images, significantly enhances structural integrity, texture clarity, and visual naturalness, and suppresses noise, artifacts, and color cast issues.

[0027] 5. Color Refinement Mechanism Based on Joint Frequency and Spatial Domain Modeling: A Color Refinement Mechanism (CRM) is constructed using an encoder-decoder structure. Fast Fourier Representation is introduced at the bottleneck layer to achieve joint modeling of spatial domain texture information and frequency domain global information. This is used for global color calibration and local detail optimization of the fusion results. Noise, color cast, and artifacts are eliminated with only a few additional parameters.

[0028] 6. Lightweight Network Structure Design: The overall framework employs lightweight design strategies such as depthwise separable convolutions and dense connections. In the feature extraction stages of AMEA and MEF, depthwise separable convolutions replace some standard convolutions to reduce computational load, and dense connections improve feature reuse efficiency and reduce redundant computation. This reduces the number of model parameters and computational complexity while maintaining enhanced quality, making it suitable for edge device deployment. The overall network structure avoids complex iterative processes or large-scale pyramid structures, features a simple inference path, and is computationally stable. It can run in real-time on edge devices with limited computing power, achieving an effective balance between enhanced performance, model lightweighting, and deployment feasibility.

[0029] 7. End-to-End Collaborative Optimization System: An end-to-end collaborative optimization system is constructed, comprising AMEA, MEF, and CRM modules. This system forms a hierarchical, progressive modeling workflow, from exposure generation and multi-exposure fusion to fine-grained correction. AMEA provides diverse lighting and color priors, MEF achieves fine-grained perceptual fusion of multi-exposure information, and CRM completes global color and local structure refinement. All three modules are jointly trained under a unified objective constraint, achieving step-by-step optimization from global lighting distribution modeling to local texture restoration. Compared to existing technologies where each stage is designed independently and lacks collaboration, this system maintains stable enhancement effects in complex low-light scenes such as extremely dark conditions, high dynamic range, and severely uneven lighting, significantly improving the model's robustness and generalization ability. Attached Figure Description

[0030] Figure 1 This is the overall structure of MEGF-Net of the present invention.

[0031] Figure 2 These are candidate images generated by the adaptive multi-exposure adjustment module of this invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] This invention proposes a lightweight low-light image enhancement framework, MEGF-Net, such as... Figure 1As shown, this framework achieves performance comparable to or even better than existing non-lightweight methods while significantly reducing model complexity, especially maintaining high fidelity in detail preservation and high dynamic range control. The core of the framework consists of two main modules: AMEA (Adaptive Multi-Exposure Adjustment Module) and MEF (Multi-Exposure Fusion Module), supplemented by CRM (Color Refinement Module) for post-processing. The AMEA module adaptively predicts a set of gamma coefficients, linear modulation factors, and compensation parameters based on the input image, generating a set of candidate images with significant differences in brightness, contrast, and color response. This provides multi-dimensional exposure and rich color priors for subsequent fusion, significantly enhancing the model's adaptability to complex lighting and high dynamic range scenes. Subsequently, the MEF module adaptively fuses the candidate images and, combined with a perceptual consistency compensation mechanism, effectively maintains structural integrity, texture clarity, and local color naturalness, achieving an optimized balance between visual naturalness and detail preservation. Furthermore, inspired by the latest frequency domain enhancement methods, the CRM module adopts an encoder-decoder structure and introduces Fast Fourier Representation at the bottleneck layer to jointly capture local spatial texture and global frequency information. This enables global color calibration and local detail refinement of the fused image, thereby further improving visual consistency and naturalness. The following sections will detail MEGF-Net's two core components, AMEA and MEF.

[0034] 1. Adaptive Multi-Exposure Adjustment Module The AMEA module proposed in this invention is inspired by Zero-DCE. Zero-DCE transforms the low-light image enhancement problem into an estimation problem of higher-order curves for pixel values. By performing nonlinear mapping on each pixel and iteratively optimizing, it achieves effective enhancement of low-light images. However, this method has significant limitations. First, performing nonlinear mapping on pixel values ​​alone ignores the global and local brightness information of the image. Second, iterative optimization is prone to error accumulation, especially in high dynamic range scenes, where the enhancement effect may be unsatisfactory. Nevertheless, the higher-order curve estimation method proposed by Zero-DCE provides important insights for low-light image enhancement. Based on this idea, we designed the AMEA module, which inherits the advantages of nonlinear mapping and effectively alleviates the problems of missing global brightness and accumulation of iterative errors through a joint prediction mechanism.

[0035] AMEA's core innovation lies in jointly predicting gamma coefficients, linear modulation factors, and compensation parameters to generate a set of candidate images with significant differences in brightness, contrast, and color response. This design not only achieves non-linear brightness adjustment of the entire image but also allows for linear modulation and overall offsetting of each pixel, forming a joint model of the illumination and color distribution of low-light images. It is worth emphasizing that AMEA is not simply a fixed brightness scaling or exposure adjustment; rather, it uses joint learning to form an image set with multi-view exposure interpretation capabilities, providing rich illumination and color priors for subsequent fusion modules, thereby significantly improving the model's adaptability and enhancement effects in complex low-light environments. The overall architecture of AMEA is as follows: Figure 1 As shown.

[0036] In terms of specific implementation, firstly, a 3×3 convolution is used to extract the low-light image. Extracting shallow features This serves as the feature basis for subsequent deep feature extraction and prediction of gamma coefficients, linear modulation factors, and compensation parameters. Deep feature extraction is performed by the FEU (Feature Extraction Unit), which employs a densely connected cascaded convolutional structure: (1) in, It consists of four cascaded convolutions. To achieve lightweight design, the first three convolutions use depthwise convolutions, and the last convolution uses a standard 3×3 convolution. At the same time, dense connections ensure efficient use of features. (2) in Indicates channel cascading. For the first i The output of the depthwise convolution, the... i The input to each depthwise convolution is .

[0037] Then, The inputs are fed into the GAU (Global Attention Unit) and SAU (Spatial Attention Unit) for calculating the gamma coefficients and linear modulation factor, respectively. The compensation parameters are estimated based on the GAU and the linear modulation factor. (3) (4) (5) in, N To estimate the number of parameter sets, in this invention, N =6; , and These represent the gamma coefficients, linear modulation factor, and compensation parameters estimated under the i-th set of parameters, respectively. and These represent GAU and SAU operations, respectively.

[0038] like Figure 1 As shown, SAU first uses 1×1 convolutions to increase the number of feature channels from C to 4C, then processes single-channel local features through depthwise convolutions, and finally reduces the number of feature channels to 3N using 1×1 convolutions. This design allows features to be learned simultaneously in both channel and spatial dimensions, achieving attention to both location and local texture.

[0039] In the GAU processing, GAU first obtains global information through adaptive average pooling, and then maps deep features through two linear layers to obtain gamma coefficients and compensation parameters. In this process, the data channel dimension first increases from C to 4C, and then decreases to 3N to achieve efficient learning of global parameters.

[0040] In obtaining , and Then, AMEA performs a non-linear transformation on the low-light image to generate a set of candidate images: (6) This represents the image under the i-th set of parameters, i.e., candidate images corresponding to different illumination and color distributions, providing diverse illumination and color prior information for subsequent MEF modules. The estimated... N Candidate images such as Figure 2 As shown, Figure 2 (a) and Figure 2 (h) represent the input low-light image and the ground truth image, respectively. Figure 2 (b) to Figure 2 (g) Six candidate images generated by the AMEA module. It is evident that each candidate image exhibits significant differences in brightness, contrast, and color performance, validating the effectiveness of the proposed method in generating diverse exposure representations. By providing a set of candidate images with diverse lighting, texture, and color priors, the AMEA module provides sufficient features for subsequent MEF fusion, effectively improving the naturalness and detail of the enhanced images.

[0041] 2. Multi-exposure fusion module To fully leverage the brightness and color priors in the multi-exposure candidate images generated by the AMEA module and obtain enhanced results with higher structural consistency and visual fidelity in real-world scenes, we propose an adaptive MEF module. This module performs fine-grained weighting on candidate images at the channel level and further introduces a perceptual consistency compensation mechanism, making the final output more consistent with the human visual system's perceptual preferences for illumination, color, and local structure.

[0042] First, to enhance the model's sensitivity to the inherent structure and color distribution of low-light scenes, we concatenate the input low-light image and multi-exposure candidate images along the channel dimension. This means that the features contain N+1 images. This design explicitly incorporates the original low-light image as a structural constraint into the fusion process, so that the model avoids deviating from the basic structure of the real scene when adjusting brightness and color.

[0043] Subsequently, 3×3 convolution and FEU were used to extract deep features to characterize the brightness differences, local texture changes, and color distribution patterns between multi-exposure images: (7) Based on this, GAU is used to predict the fusion coefficients channel by channel. Since N+1 images need to be fused, the input features of GAU... The number of channels is C, and the number of channels of the output fusion coefficient features is 3N+3, realizing fine-grained adaptive weighting of the R, G, and B channels of the candidate image: (8) Compared to traditional global weighting or fixed fusion strategies, this design can simultaneously model local exposure differences in both spatial and channel dimensions, significantly enhancing details in both bright and dark areas. However, channel-wise fusion coefficients primarily focus on brightness and texture information, lacking explicit modeling of human visual perception preferences, which may lead to subtle deviations in color distribution or brightness style from the real scene in the fusion result. Therefore, we further refine the low-light image with features... The images are stitched together, and the perceptual parameters are estimated using GAU. Since the perceptual compensation strategy refines the final fused image, the number of channels for the GAU input features is C+3, and the number of output features is 3.

[0044] (9) This perceptual compensation strategy enables fine-grained correction of the fused brightness and color, resulting in a more visually natural, stable, and consistent lighting style in the output image. Finally, the MEF module generates the final enhanced image based on the fusion coefficients and perceptual parameters. The results for the R, G, and B channels of the final enhanced image are as follows: (10) (11) (12) in, express The value of the m-th channel. .

[0045] Overall, the MEF module achieves systematic innovation in multi-exposure information modeling, channel-level adaptive fusion, and visual perception compensation. It not only fully leverages the rich brightness and color priors in multi-exposure candidate images, but also enhances the visual credibility of the augmented results in terms of structural, brightness, and color consistency by explicitly introducing a perception compensation strategy.

[0046] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-light image enhancement system based on multi-exposure generation and fusion, characterized in that, include: An adaptive multi-exposure adjustment module is used to jointly predict the gamma coefficient, linear modulation factor and compensation parameters based on the input low-light image, and perform nonlinear transformation on the low-light image to generate multiple multi-exposure candidate images with different brightness, contrast and color response. The multi-exposure fusion module is used to perform channel stitching of the low-light image and the multi-exposure candidate image, and to calculate the fusion feature image; The channel-by-channel fusion coefficient is calculated based on the fusion feature image, and the perceptual compensation parameter is calculated based on the stitched image of the low-light image and the fusion feature image. The color refinement module performs fine-grained weighted fusion of the multi-exposure candidate images using the channel-by-channel fusion coefficients, and corrects the fusion result using perceptual compensation parameters to obtain the final enhanced image.

2. The system according to claim 1, characterized in that, The adaptive multi-exposure adjustment module includes: The feature extraction unit is used to extract shallow features from the low-light image and extract deep features through a densely connected cascaded convolutional structure. Spatial attention unit (SAU) is used to predict the linear modulation factor based on the deep features. ; Global attention unit (GAU) is used to predict gamma coefficients based on the deep features. Based on linear modulation factor Predicted compensation parameters ; The adaptive multi-exposure adjustment module generates the multi-exposure candidate image according to the following formula: in, For low-light images, This represents the candidate image under the i-th set of parameters, that is, the candidate image corresponding to different brightness, contrast and color response.

3. The system according to claim 2, characterized in that, The feature extraction unit adopts a depthwise separable convolution and dense connection structure: the first three convolutions are depthwise convolutions, the last convolution is a standard convolution, and the input of each convolution is the channel concatenation of the outputs of all previous convolutions and the shallow features.

4. The system according to claim 2, characterized in that, The global attention unit obtains global information through adaptive average pooling and performs feature mapping through two linear layers in sequence, increasing the channel dimension first and then decreasing it to the output dimension; the spatial attention unit increases the number of channels through 1×1 convolution, processes it through depthwise convolution, and then decreases the number of channels to the output dimension through 1×1 convolution.

5. The system according to claim 1, characterized in that, The process by which the multi-exposure fusion module obtains the fused feature image includes: the feature extraction unit extracts deep features from the stitched image, and then convolves them into the fused feature image using a 3×3 convolution.

6. The system according to claim 1, characterized in that, Predicting channel-by-channel fusion coefficients using Global Attention Unit (GAU) , , N To estimate the number of parameter groups.

7. The system according to claim 6, characterized in that, The perceptual compensation parameters are calculated using a Global Attention Unit (GAU). .

8. The system according to claim 7, characterized in that, The color refinement module yields the following results for the R, G, and B channels of the final enhanced image: (10) (11) (12) in, express The value of the m-th channel. .

9. The system according to any one of claims 2-8, characterized in that, The Spatial Attention Unit (SAU) first uses 1×1 convolutions to increase the number of feature channels from C to 4C, then processes single-channel local features through depthwise convolutions, and finally reduces the number of feature channels to 3N through 1×1 convolutions. N To estimate the number of parameter groups.

10. The system according to any one of claims 2-8, characterized in that, The Global Attention Unit (GAU) first acquires global information through adaptive average pooling, and then maps deep features through two linear layers. In this process, the data channel dimension first increases from C to 4C, and then decreases to 3N. N To estimate the number of parameter groups.

Images