Compressed domain weak light image enhancement method and system based on 2D Gaussian splash
By directly enhancing the image within the compression domain, and utilizing the 2D Gaussian splashing compression domain method, the problems of low computational efficiency and secondary distortion in traditional methods are solved, achieving efficient and natural low-light image enhancement effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-01-21
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies suffer from low computational efficiency and secondary distortion when processing compressed images based on 2D Gaussian splashing. In particular, during the enhancement of low-light images, traditional methods require a cumbersome decompression-enhancement-recompression process, resulting in computational redundancy and a decline in image quality.
We propose a compression domain low-light image enhancement method based on 2D Gaussian splashing. By constructing an image primitive representation model, we optimize the photometric attributes of Gaussian primitives using a semantically guided hybrid expert enhancement network. The method is trained using a multi-objective unsupervised loss function, and the enhancement is completed directly in the compression domain, avoiding the decompression and recompression processes.
It achieves efficient image enhancement while maintaining a high compression rate, avoiding computational redundancy and secondary damage. It has zero-sample unsupervised learning capability, adapts to various low-light image scenarios, and improves the naturalness and realism of images.
Smart Images
Figure CN121961955A_ABST
Abstract
Description
A Compressed Domain Low-Light Image Enhancement Method and System Based on 2D Gaussian Splashing Technical Field
[0001] This invention belongs to the field of image processing and computer vision technology, and relates to a compressed domain low-light image enhancement method and system based on 2D Gaussian splashing. Background Technology
[0002] Image signals are inherently continuous and complex, and for a long time, computer vision tasks have mainly relied on discrete pixel grids as the underlying representation. With the pursuit of high fidelity and extreme compression ratio, 2D Gaussian Splatting (2DGS), as an emerging explicit scene representation, has shown the potential to surpass traditional methods due to its high compression ratio and real-time rendering characteristics.
[0003] However, in practical applications such as autonomous driving and security monitoring, images often appear in low-light conditions due to underexposure. Existing Low Light Image Enhancement (LLIE) algorithms, whether traditional histogram equalization or deep learning-based methods (such as Retinex-Net and EnlightenGAN), are all built on the pixel domain. This means that when processing 2DGS-based compressed images, a cumbersome process of "2DGS rendering and decompression - pixel domain enhancement - 2DGS recompression" must be followed.
[0004] This path dependence leads to serious efficiency and quality bottlenecks: First, the repeated rendering and optimization processes introduce high computational redundancy, undermining the real-time advantage of 2DGS; second, noise and artifacts introduced by pixel-domain enhancement are amplified during secondary compression, resulting in secondary damage to image information. Therefore, how to achieve illumination enhancement directly within the compressed representation domain while maintaining the high compression ratio of 2DGS is a pressing technical challenge that needs to be addressed. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a compressed domain low-light image enhancement method and system based on 2D Gaussian splashing, named LL-GaussianImage, to solve the problems of low computational efficiency and secondary distortion in the existing technology for compressed image enhancement.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A compressed-domain low-light image enhancement method based on 2D Gaussian splashing is proposed, comprising the following steps: S1, constructing an image primitive representation model based on 2D Gaussian splashing, transforming the input low-light image from the pixel domain into an explicit compressed representation composed of a set of discrete 2D Gaussian primitives, and optimizing and reconstructing the geometric properties and initial luminance properties of the Gaussian primitives; S2, constructing a semantically guided hybrid expert enhancement network model, which includes a semantically perceptual feature extraction module, a spatially adaptive weight estimation module, and a hybrid expert color transformation module based on residual MLP; S3, freezing the geometric properties of the optimized and reconstructed Gaussian primitives, and using the hybrid expert enhancement network model to reparameterize and optimize the luminance properties of the Gaussian primitives, constraining the training process through a multi-objective unsupervised loss function to obtain the enhanced Gaussian luminance parameters; S4, using the enhanced Gaussian luminance parameters and the frozen geometric parameters, generating the final enhanced image through differentiable rasterization rendering.
[0008] Furthermore, in step S1, the input low-light image is first... Modeled as a 2D Gaussian set:
[0009] Each Gaussian Consists of a set of learnable parameters The only certainty is that, Indicates the coordinates of the center position. Represents the 2D covariance matrix. Represents RGB color attributes. Indicates opacity; optimizes the reconstruction process by minimizing the rendered image. Compared with the original low-light image Reconstruction losses between To obtain the optimized parameter set The reconstruction loss includes L1 loss and multi-scale structural similarity loss, which is expressed as:
[0010] in, The weighting coefficients represent the weights that balance the L1 loss and the multi-scale structural similarity loss. express and The structural similarity index between them.
[0011] Furthermore, in step S2, the constructed semantic-aware feature extraction module uses a lightweight convolutional neural network as its backbone to extract multi-scale semantic features from the input image; the spatial adaptive weight estimation module maps the semantic features into a spatially aligned hybrid weight map. And for each Gaussian element, a mixed weight vector is obtained by sampling. The expert color transformation module uses weights... Calculate the enhanced color.
[0012] Furthermore, the semantic-aware feature extraction module uses low-light images As input, the first 7 layers of a pre-trained MobileNetV2 module are used as the feature extractor. This is to capture low-level texture and mid-level semantic information. The feature extraction process can be represented as:
[0013] in, For feature dimension, To achieve the downsampled resolution, a spatially aligned hybrid weight head was designed to map continuous image features onto a discrete Gaussian point cloud. This head network then maps the features... Mapping to the weight space generates a low-resolution weight graph. .in, The number of predefined color transformation operators; the spatial adaptive weight estimation module calculates the sampling weight for each Gaussian unit. The process is as follows:
[0014] in, The operation is applied to the channel dimension to ensure ; Bilinear interpolation operation; the hybrid expert color transformation module includes... A color transformation operator, wherein This is a learnable residual multilayer perceptron operator. 1 is the identity mapping operator; the 2nd is the 3rd. Color enhancement by Gaussian elements The calculation formula is as follows:
[0015] in, Original color, The weights of the identity operators, For the first The weights of a learnable operator Color transformation in residual form: , It is an MLP network.
[0016] Furthermore, in step S3, the geometric properties of the optimized and reconstructed Gaussian elements are frozen, that is, the coordinates of the center position are frozen. and 2D covariance matrix Design a multi-objective unsupervised loss function :
[0017] in, In order to expose the losses, To preserve hue, For spatial consistency loss, For color constancy loss, For histogram contrast loss, For regularization terms, The corresponding weight coefficients are used; the enhanced Gaussian photometric parameters are obtained by constraining the training process through a multi-objective unsupervised loss function.
[0018] Furthermore, in multi-objective unsupervised loss functions In the middle, exposure loss To correct for lighting levels, the average image brightness is pulled towards the target value. It is represented as:
[0019] in, Represents the luminance component of an image; hue preservation loss. In HSV space calculations, a saturation mask is introduced, which is represented as:
[0020] in, For indicator functions, The saturation threshold, and The enhanced image and the original image are respectively in pixels. The hue value at that location, Saturation mask; spatial consistency loss The Sobel operator is used to extract gradient maps, and the L1 loss is calculated to maintain the gradient distribution of the enhanced image consistent with that of the original image and suppress noise. This is expressed as follows:
[0021] Color richness loss Based on color space theory, the perceptual vividness of an image is enhanced by maximizing the statistical dispersion of these two antagonistic channels, as expressed below:
[0022] in, and These represent the images in and Spatial standard deviation on the channel This is a numerically stable term. This constraint enables the model to effectively stretch the color distribution while increasing brightness, thus avoiding a "hazy" visual appearance in the image.
[0023] Histogram contrast loss Used to limit the standard deviation of enhanced images Not lower than the standard deviation of the original image a certain proportion The standard deviation of the constrained augmented image is expressed as:
[0024] in, This is the preset contrast enhancement factor.
[0025] Furthermore, in step S4, the optimized photometric parameters and the original geometric parameters are input into a differentiable rasterizer, and the final high-quality enhanced image is synthesized according to the depth sorting and alpha mixing formula:
[0026] in, It indicates that it is the first Each Gaussian element in a pixel The instantaneous opacity at a given point is determined by its reference opacity. Determined together with the spatial attenuation term; indicates the distance the light travels to the first digit. Transmittance at the number of elements, i.e., the previous The remaining amount of light that is not blocked by each primitive.
[0027] On the other hand, a compressed-domain low-light image enhancement system based on 2D Gaussian splashing is also provided to implement the aforementioned compressed-domain low-light image enhancement method based on 2D Gaussian splashing. This system includes: an image primitive reconstruction module, a semantic-guided enhancement module, an optimization training module, and a rendering output module. The image primitive reconstruction module utilizes a 2D Gaussian splashing compression reconstruction strategy to transform the input low-light image into discrete geometric primitive representations and optimize geometric and photometric parameters. The semantic-guided enhancement module incorporates a semantically guided hybrid expert enhancement network model, including feature extraction, weight estimation, and color transformation sub-modules. The optimization training module iteratively optimizes the hybrid expert enhancement network based on a multi-objective unsupervised loss function to generate enhanced Gaussian photometric parameters. The rendering output module uses the enhanced Gaussian parameters to perform differentiable rasterization and outputs the final enhanced image.
[0028] The beneficial effects of this invention are as follows: First, compression is equivalent to enhancement. As the first framework to directly perform low-light enhancement within the 2DGS compressed representation domain, this invention cleverly avoids the computational redundancy problem caused by the cumbersome "decompression-enhancement-recompression" process of traditional methods. Traditional methods involve switching between different steps, which not only consumes a lot of computational resources but may also lead to image quality degradation due to multiple processing steps. In contrast, this invention operates directly in the compressed domain, efficiently and accurately completing the enhancement task.
[0029] Secondly, it employs zero-shot unsupervised learning. This invention does not require paired low-light and normal-light training data; it achieves enhancement solely through self-supervised iteration of a single image, demonstrating extremely strong generalization capabilities. This means that it can quickly adapt and effectively enhance any low-light image, regardless of the scene or type, without requiring significant effort to collect and label specific data.
[0030] Third, there is no secondary damage. This invention strictly guarantees the structural fidelity of the image by freezing geometric attributes and optimizing only photometric attributes, successfully avoiding artifacts and noise amplification problems that may be caused by pixel domain enhancement, making the enhanced image more natural and realistic.
[0031] Fourth, it is highly efficient and lightweight. Compared to pixel grid representation, this invention can reconstruct and enhance images using a very small number of primitives (with a compression ratio of over 50 times), significantly reducing storage and transmission costs and making it extremely valuable for applications in resource-constrained environments.
[0032] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 is a schematic diagram of the overall architecture of the compressed domain low-light image enhancement method based on 2D Gaussian splashing in an embodiment of the present invention; Figure 2 is a visualization diagram of semantic hybrid weights in an embodiment of the present invention; Figure 3 is a structural decomposition diagram of the hybrid expert enhancement module in an embodiment of the present invention; Figure 4 is a schematic diagram comparing the different representation domain enhancement schemes of the present invention with the existing SOTA pixel domain enhancement method in an embodiment of the present invention, Figure 4(a) is the SOTA pixel domain enhancement method, and Figure 4(b) is the 2DGS compressed domain scheme proposed in the present invention. Detailed Implementation
[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0035] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0036] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0037] Please refer to Figures 1 to 4, which illustrate a compressed domain low-light image enhancement method and system based on 2D Gaussian splashing.
[0038] Example 1 This example provides a compressed domain low-light image enhancement method based on 2D Gaussian splashing, as shown in Figure 1. The specific process is as follows: Step S1: 2DGS image primitive reconstruction: Construct an image primitive representation model based on 2D Gaussian splashing, transforming the input low-light image from the pixel domain into an explicit compressed representation composed of a set of discrete 2D Gaussian primitives, and optimizing and reconstructing the geometric properties and initial luminance properties of the Gaussian primitives; First, the input low-light image Modeled as a 2D Gaussian metaset Each primitive is composed of positions. Covariance ,color and opacity Definition. Image edges are fitted using an anisotropic Gaussian distribution, and the loss function is minimized. To iteratively optimize parameters loss function Represented as:
[0039] In this embodiment, the number of iterations was set to 30,000, and efficient encoding of continuous image signals was finally achieved using very few primitives.
[0040] Step S2: Construct a semantically guided hybrid expert augmentation network, which includes a semantically aware feature extraction module, a spatially adaptive weight estimation module, and a hybrid expert color transformation module based on residual MLP.
[0041] The semantic-aware feature extraction module uses a lightweight convolutional neural network as its backbone to extract multi-scale semantic features from the input image. In this embodiment, a lightweight CNN is designed as a feature extractor to extract multi-scale semantic features from the input image. As input, the first 7 layers of a pre-trained MobileNetV2 module are used as the feature extractor. This is to capture low-level texture and mid-level semantic information. The feature extraction process can be represented as:
[0042] in, For feature dimension, To achieve the downsampled resolution, a spatially aligned hybrid weight head was designed to map continuous image features onto a discrete Gaussian point cloud. This head network then maps the features... Mapping to the weight space generates a low-resolution weight graph. .in, It is the number of predefined color transformation operators; the spatial adaptive weight estimation module maps semantic features to spatially aligned hybrid weight graphs. And for each Gaussian element, a mixed weight vector is obtained by sampling. In this embodiment, as shown in Figure 2, a head network maps features to a hybrid weight map. The spatial adaptive weight estimation module calculates the sampling weights for each Gaussian unit. The process is as follows:
[0043] in, The operation is applied to the channel dimension to ensure ; Bilinear interpolation operation; included in the Mixing Expert Color Transform module. A color transformation operator, wherein Each is a learnable residual multilayer perceptron (MLP) operator. 1 is the identity mapping operator; the 2nd is the 3rd. Color enhancement by Gaussian elements The calculation formula is as follows:
[0044] in, Original color, The weights of the identity operators, For the first The weights of a learnable operator Color transformation in residual form: , It is an MLP network.
[0045] In this embodiment, as shown in Figure 3, the construction is as follows: The system employs a color transformation operator, including one identity operator and 15 learnable operators based on residual MLPs. Each MLP consists of three fully connected layers, utilizing residual connections to ensure the stability of the initial state. Bilinear interpolation is used in the weight graph. The above are the sampling weights for each Gaussian element. And calculate the enhanced color according to the formula.
[0046] Step S3: Zero-shot unsupervised optimization: Freeze the geometric attributes obtained in step S1, and use a hybrid expert augmentation network model to reparameterize and optimize the photometric attributes of Gaussian units. The training process is constrained by a multi-objective unsupervised loss function to obtain the augmented Gaussian photometric parameters; freeze the positions learned in S1. Covariance Only optimize and enhance network parameters. Define the total loss function. :
[0047] in, In order to expose the losses, To preserve hue, For spatial consistency loss, For color constancy loss, For histogram contrast loss, For regularization terms, These are the corresponding weighting coefficients.
[0048] Specifically, To compensate for exposure loss and correct for lighting levels, the average image brightness is pulled towards the target value. It is represented as:
[0049] in, Represents the luminance component of an image; hue preservation loss. In HSV color space calculations, a saturation mask is introduced to prevent hue calculation errors in low-saturation areas (gray / white) from interfering with training, ensuring that the enhanced colors are not distorted.
[0050] in, For indicator functions, The saturation threshold, and The enhanced image and the original image are respectively in pixels. The hue value at that location, Saturation mask; spatial consistency loss The Sobel operator is used to extract gradient maps, and the L1 loss is calculated to maintain the gradient distribution of the enhanced image consistent with that of the original image and suppress noise. This is expressed as follows:
[0051] Color richness loss Based on color space theory, the perceptual vividness of an image is enhanced by maximizing the statistical dispersion of these two antagonistic channels, as expressed below:
[0052] in, and These represent the images in and Spatial standard deviation on the channel This is a numerically stable term. This constraint enables the model to effectively stretch the color distribution while increasing brightness, thus avoiding a "hazy" visual appearance in the image.
[0053] Histogram contrast loss Used to limit the standard deviation of enhanced images Not lower than the standard deviation of the original image a certain proportion The standard deviation of the constrained augmented image is expressed as:
[0054] in, This is the preset contrast enhancement factor.
[0055] Based on the total loss function The training iterations were set to 50,000, and the Adam optimizer was used, with the learning rate decaying in conjunction with a cosine annealing strategy.
[0056] Step S4: Rendering Output: Using the enhanced Gaussian photometric parameters and frozen geometric parameters, the final enhanced image is generated through differentiable rasterization rendering. The optimized photometric parameters and original geometric parameters are input into the differentiable rasterizer, and the final high-quality enhanced image is synthesized according to depth sorting and the alpha blending formula.
[0057] in, It indicates that it is the first Each Gaussian element in a pixel The instantaneous opacity at a given point is determined by its reference opacity. Determined together with the spatial attenuation term; indicates the distance the light travels to the first digit. Transmittance at the number of elements, i.e., the previous The remaining amount of light that is not blocked by each primitive.
[0058] Example 2 This example provides a compressed domain low-light image enhancement system based on 2D Gaussian splashing, which is used to execute the aforementioned compressed domain low-light image enhancement method based on 2D Gaussian splashing. The system includes an image primitive reconstruction module, a semantic-guided enhancement module, an optimization training module, and a rendering output module. The image primitive reconstruction module uses a 2D Gaussian splashing compression reconstruction strategy to transform the input low-light image into discrete geometric primitive representations and optimizes geometric and photometric parameters. The semantic-guided enhancement module incorporates a semantically guided hybrid expert enhancement network model, including feature extraction, weight estimation, and color transformation sub-modules. The optimization training module iteratively optimizes the hybrid expert enhancement network based on a multi-objective unsupervised loss function to generate enhanced Gaussian photometric parameters. The rendering output module uses the enhanced Gaussian parameters to perform differentiable rasterization and outputs the final enhanced image.
[0059] Based on the methods and systems in Embodiments 1 and 2, this embodiment compares the proposed solution with existing state-of-the-art (SOTA) pixel domain enhancement methods. Figure 4 shows a schematic diagram of the comparison results for different representation domain enhancement schemes. Figure 4(a) shows the SOTA pixel domain enhancement method, and Figure 4(b) shows the 2DGS compression domain scheme proposed in this invention. Experimental results show that on the LOL and LSRW datasets, this method maintains a high compression ratio (approximately 50 times), while achieving PSNR and SSIM indices that are superior to or comparable to existing SOTA pixel domain enhancement methods. Furthermore, the visual effect is natural in color and without obvious artifacts.
[0060] Finally, it should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A compressed domain low-light image enhancement method based on 2D Gaussian splashing, characterized in that: The method includes the following steps: S1, constructing an image primitive representation model based on 2D Gaussian splashing, transforming the input low-light image from the pixel domain into an explicit compressed representation composed of a set of discrete 2D Gaussian primitives, and optimizing and reconstructing the geometric properties and initial luminance properties of the Gaussian primitives; S2, constructing a semantically guided hybrid expert enhancement network model, which includes a semantically perceptual feature extraction module, a spatially adaptive weight estimation module, and a hybrid expert color transformation module based on residual MLP; S3, freezing the geometric properties of the optimized and reconstructed Gaussian primitives, using the hybrid expert enhancement network model to reparameterize and optimize the luminance properties of the Gaussian primitives, and obtaining the enhanced Gaussian luminance parameters by constraining the training process through a multi-objective unsupervised loss function; S4, using the enhanced Gaussian luminance parameters and the frozen geometric parameters, generating the final enhanced image through differentiable rasterization rendering.
2. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 1, characterized in that: In step S1, the input low-light image is first... Modeled as a 2D Gaussian set: Each Gaussian Consists of a set of learnable parameters The only certainty is that, Indicates the coordinates of the center position. Represents the 2D covariance matrix. Represents RGB color attributes. Indicates opacity; optimizes the reconstruction process by minimizing the rendered image. Compared with the original low-light image Reconstruction losses between To obtain the optimized parameter set The reconstruction loss includes L1 loss and multi-scale structural similarity loss, which is expressed as: in, The weighting coefficients represent the balance between L1 loss and multi-scale structural similarity loss. express and The structural similarity index between them.
3. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 2, characterized in that: In step S2, the constructed semantic-aware feature extraction module uses a lightweight convolutional neural network as its backbone to extract multi-scale semantic features from the input image; the spatial adaptive weight estimation module maps the semantic features into a spatially aligned hybrid weight map. And for each Gaussian element, a mixed weight vector is obtained by sampling. The expert color transformation module uses weights... Calculate the enhanced color.
4. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 3, characterized in that: The semantic-aware feature extraction module uses low-light images As input, the first 7 layers of a pre-trained MobileNetV2 module are used as the feature extractor. To capture low-level texture and mid-level semantic information, the feature extraction process is represented as follows: in, For feature dimension, To achieve the downsampled resolution, a spatially aligned hybrid weight head is designed to map continuous image features onto a discrete Gaussian point cloud. This head network then... Mapping to the weight space generates a low-resolution weight graph. ,in, The number of predefined color transformation operators; the spatial adaptive weight estimation module calculates the sampling weight for each Gaussian unit. The process is as follows: in, The operation is applied to the channel dimension to ensure ； Bilinear interpolation operation; the hybrid expert color transformation module includes... A color transformation operator, wherein This is a learnable residual multilayer perceptron operator. 1 is the identity mapping operator; the 2nd is the 3rd. Color enhancement by Gaussian elements The calculation formula is as follows: in, Original color, The weights of the identity operators, For the first The weights of a learnable operator Color transformation in residual form: ， It is an MLP network.
5. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 3, characterized in that: In step S3, the geometric properties of the optimized and reconstructed Gaussian elements are frozen, that is, the coordinates of the frozen center position are frozen. and 2D covariance matrix Design a multi-objective unsupervised loss function ： in, In order to expose the losses, To preserve hue, For spatial consistency loss, For color constancy loss, For histogram contrast loss, For regularization terms, The corresponding weight coefficients are used; the enhanced Gaussian photometric parameters are obtained by constraining the training process through a multi-objective unsupervised loss function.
6. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 5, characterized in that: In multi-objective unsupervised loss function In the middle, exposure loss To correct for lighting levels, the average image brightness is pulled towards the target value. It is represented as: in, Represents the luminance component of an image; hue preservation loss. In HSV space calculations, a saturation mask is introduced, which is represented as: in, For indicator functions, The saturation threshold, and The enhanced image and the original image are respectively in pixels. The hue value at that location, Saturation mask; spatial consistency loss The Sobel operator is used to extract gradient maps, and the L1 loss is calculated to maintain the gradient distribution of the enhanced image consistent with that of the original image and suppress noise. This is expressed as follows: Color richness loss Based on color space theory, the perceptual vividness of an image is enhanced by maximizing the statistical dispersion of these two antagonistic channels, as expressed below: in, and These represent the images in and Spatial standard deviation on the channel Numerical stability term; histogram contrast loss Used to limit the standard deviation of enhanced images Not lower than the standard deviation of the original image a certain proportion The standard deviation of the constrained augmented image is expressed as: in, This is the preset contrast enhancement factor.
7. The compressed domain low-light image enhancement method based on 2D Gaussian splashing according to claim 5, characterized in that: In step S4, the optimized photometric parameters and the original geometric parameters are input into a differentiable rasterizer, and the final high-quality enhanced image is synthesized according to the depth sorting and alpha mixing formula. in, It indicates that it is the first Each Gaussian element in a pixel The instantaneous opacity at a given point is determined by its reference opacity. Determined together with the spatial attenuation term; indicates the distance the light travels to the first digit. Transmittance at the number of elements, i.e., the previous The remaining amount of light that is not blocked by each primitive.
8. A compressed domain low-light image enhancement system based on 2D Gaussian splashing, characterized in that: The method for enhancing low-light images in the compressed domain based on 2D Gaussian splashing as described in any one of claims 1-7 includes: an image primitive reconstruction module, a semantically guided enhancement module, an optimization training module, and a rendering output module. The image primitive reconstruction module utilizes a 2D Gaussian splashing compression reconstruction strategy to transform the input low-light image into discrete geometric primitive representations and optimizes geometric and photometric parameters. The semantically guided enhancement module incorporates a semantically guided hybrid expert enhancement network model, including feature extraction, weight estimation, and color transformation sub-modules. The optimization training module iteratively optimizes the hybrid expert enhancement network based on a multi-objective unsupervised loss function to generate enhanced Gaussian photometric parameters. The rendering output module performs differentiable rasterization using the enhanced Gaussian parameters and outputs the final enhanced image.