A Low-Light Image Enhancement Method Integrating Continuous Gaussian Rendering and Retinex

Abstract

This invention discloses a low-light image enhancement method integrating continuous Gaussian rendering and Retinex, belonging to the field of computer vision and image processing. The method first performs hierarchical feature encoding on the input low-light image, extracting deep contextual features and shallow high-frequency jump features. It then generates two sets of continuous Gaussian primitive parameters—one for illumination and one for features—through a dual-branch prediction head. These parameters are then passed through a continuous Gaussian renderer with learnable temperature to obtain a spatially continuous global illumination field and contextual feature field. A basic reflectance map is obtained based on Retinex inverse operations, and a coordinate-driven implicit neural network is used to predict the reflectance residual map. The two are then fused to output the enhanced image. This invention fundamentally eliminates the mesh artifacts of discrete Gaussian modeling, achieving an optimal balance between illumination smoothness, color fidelity, and detail recovery capability. It can be widely applied to low-light image processing in scenarios such as autonomous driving, security monitoring, and mobile photography.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing, and relates to a low-light image enhancement method that integrates continuous Gaussian rendering and Retinex. Background Technology

[0002] Low-light image enhancement is a core, low-level visual task in computer vision. As a crucial preprocessing step for high-level visual tasks such as autonomous driving, all-weather security monitoring, nighttime imaging, and smart terminal photography, its enhancement effect directly determines the recognition accuracy and reliability of downstream tasks and application scenarios. Decoupled enhancement methods based on Retinex theory have become the mainstream technical approach due to their excellent physical interpretability. These methods decompose the observed image into illumination and reflectivity components for separate optimization. However, these methods all use pixel-level discrete mesh modeling, which cannot balance global illumination smoothing and local high-frequency detail restoration when dealing with low-light areas with drastic brightness changes. This easily leads to problems such as color distortion, local overexposure, and structural jagged artifacts. At the same time, discretization modeling directly violates the physical prior of illumination smoothing in Retinex theory, resulting in extremely poor detail restoration in dark areas.

[0003] To overcome the limitations of discrete representations, implicit neural representations have been introduced into low-light image enhancement tasks. By capturing the intrinsic structure of images through coordinate-driven continuous function mapping, they alleviate the discrete artifact problem to some extent. However, pure implicit networks exhibit significant spectral bias when fitting the highly nonlinear high-frequency textures in low-light images and lack explicit geometric control. This makes it difficult to accurately reproduce severely damaged edges and texture details in low-light environments, leading to problems such as over-smoothing of details, blurred textures, and distorted color saturation, failing to meet the practical requirements of high-fidelity enhancement. Explicit image representation techniques, such as two-dimensional Gaussian scattering (2DGS), have been attempted for application in the field of low-light enhancement due to the efficient representation advantages of Gaussian units. However, existing 2DGS schemes all use a discrete sputtering mechanism, which generates pixels by querying nearest neighbor primitives or weighted aggregation of discrete points. This discrete sampling method introduces grid-like discontinuous artifacts in continuous illumination fields, directly violating the physical constraints of illumination smoothing in Retinex theory. At the same time, such techniques have not achieved deep integration with Retinex theory and cannot achieve accurate decoupling of illumination and reflectivity, further aggravating the image distortion problem.

[0004] In summary, existing low-light image enhancement technologies suffer from core bottlenecks that are difficult to address in a coordinated manner: discrete modeling leads to mesh artifacts, pure implicit representations are insufficient for high-frequency detail recovery, discrete Gaussian scattering conflicts with Retinex physical priors, and cannot simultaneously achieve global illumination smoothing, local high-frequency detail preservation, color distortion-free restoration, and noise suppression in dark areas. Furthermore, they lack a unified paradigm that combines explicit and implicit modeling, and the enhancement effects, physical rationality, and generalization capabilities in complex low-light scenes cannot meet the needs of practical applications.

[0005] To address the aforementioned shortcomings of existing technologies, this invention proposes a low-light image enhancement method that integrates continuous Gaussian rendering and Retinex. This method aims to solve the technical problems of mesh artifacts, uneven lighting distribution, loss of high-frequency texture details, color shift, and amplification of noise in dark areas that exist in existing technologies. It achieves precise decoupling of illumination and reflectivity, and significantly improves the fidelity, naturalness, and physical rationality of low-light image enhancement. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a low-light image enhancement method that integrates continuous Gaussian rendering and Retinex, aiming to overcome the technical bottleneck of existing methods that struggle to balance global illumination smoothing and local high-frequency detail restoration. By using an explicit-implicit joint modeling architecture, it achieves precise decoupling of illumination and reflectivity, ensuring the physical rationality of illumination estimation while realizing high-fidelity restoration of texture and color in dark areas. This method is suitable for the enhancement needs of various complex low-light scenes and provides stable and reliable low-light image preprocessing capabilities for downstream visual tasks such as autonomous driving, security monitoring, and mobile photography.

[0007] To achieve the above objectives, this invention provides a low-light image enhancement method that integrates continuous Gaussian rendering and Retinex, comprising the following steps: S1: Perform hierarchical feature encoding on the input low-light image to extract deep contextual features and shallow high-frequency jump features; S2: Input the deep context features into the illumination prediction head and the feature prediction head to generate a continuous Gaussian parameter set for illumination and a continuous Gaussian parameter set for features; S3: By rendering the lighting continuous Gaussian meta-parameter set and the feature continuous Gaussian meta-parameter set separately using a continuous Gaussian renderer with learnable temperature, a spatially continuous global illumination field and context feature field are obtained. S4: Based on Retinex inverse operation, the basic reflectance map is calculated from the global illumination field and the input low-light image; S5: Input the context feature field, shallow high-frequency jump features and spatial coordinate encoding into the coordinates to drive the implicit neural network and predict the reflectance residual map; S6: Linearly superimpose the base reflectance map and the reflectance residual map, and output the final enhanced image after truncation of the value range.

[0008] Optionally, the hierarchical feature encoder in step S1 adopts a dense residual connection structure, which simultaneously extracts global contextual semantic features and local high-frequency detail features containing edge and texture information through multi-level downsampling and residual skip connections.

[0009] Optionally, the illumination prediction head and the feature prediction head in step S2 are a dual-branch parallel structure. The illumination prediction head is used to model the low-frequency continuous illumination parameter field, and the feature prediction head is used to model the high-frequency texture feature parameter field.

[0010] The optional continuous Gaussian element parameter set in step S2 includes continuous center coordinates, inverse covariance matrix, opacity coefficient, and attribute parameters; wherein the inverse covariance matrix is ​​calculated by jointly using anisotropic scaling factor and rotation angle, and is used to characterize the spatial influence range and anisotropic shape of the Gaussian element.

[0011] Optionally, the continuous Gaussian renderer described in step S3 extracts K×K neighborhood Gaussian elements for continuous query coordinates, calculates Mahalanobis distance and generates exponentially decaying weights, and completes the neighborhood Gaussian attribute weighted integral, where K is a natural number.

[0012] Optionally, the learnable temperature coefficient mentioned in step S3 is mapped to a positive adaptive temperature value via the Softplus function to balance the smoothness and edge sharpness of the rendering result.

[0013] Furthermore, step S3 also includes performing an adaptive truncation operation on the global illumination field, constraining the minimum boundary of the illumination value to 0.05, to avoid numerical explosion and color bleaching caused by division operations.

[0014] Optionally, the coordinate-driven implicit neural network described in step S5 first maps the two-dimensional spatial coordinates to a high-dimensional frequency space through a random Fourier basis to capture high-frequency texture information. Then, it fuses the global semantic information of the context feature field with the local detail information of the shallow high-frequency jump features. The reflectance residual value corresponding to the coordinate position is output by the multilayer perceptron, and finally a complete reflectance residual map is generated.

[0015] Optionally, step S6 uses the additive residual Retinex paradigm to linearly fuse the base reflectance and the reflectance residual, and then uses the Clamp function to constrain the output within the [0,1] effective color space.

[0016] Furthermore, step S7 is included: Network training: The network training is completed by constructing a joint loss function based on multi-dimensional physical constraints. The joint loss function includes Charbonnier reconstruction loss, SSIM loss, VGG perceptual loss, illumination total variation loss, grayscale physical consistency loss and color cosine similarity loss, with weights of 1.0, 0.5, 0.05, 0.05, 0.05 and 0.05 for each item.

[0017] The beneficial effects of this invention are as follows: First, existing discrete Gaussian methods rely on discrete primitive sampling, which easily introduces discontinuous mesh artifacts into the illumination field, violating the physical prior of spatial smoothness in illumination. The continuous Gaussian renderer proposed in this invention models Gaussian primitives as a spatially continuous parameter field, extracts a weighted integral of neighborhood Gaussian primitives for any query coordinate, and introduces a learnable temperature coefficient to adaptively balance smoothness and edge sharpness. This mathematically guarantees a smooth transition in illumination distribution, effectively eliminating the blocky artifacts and edge tearing problems common in traditional methods, and significantly improving the physical accuracy of illumination estimation.

[0018] Secondly, existing Retinex methods either lack physical constraints leading to color distortion or couple high-frequency details with noise, resulting in texture blurring. This invention adopts an additive residual Retinex paradigm: first, the base reflectivity is calculated using a smooth lighting field obtained through continuous Gaussian rendering to ensure the physical rationality of the global color distribution; then, a coordinate-driven implicit neural network is used to predict the reflectivity residual, and random Fourier coordinate encoding is used to capture high-frequency textures. By fusing shallow jump features, accurate compensation for damaged details is achieved. This not only enables high-fidelity global color restoration but also clearly restores high-frequency information such as edges and textures in extremely dark areas, while significantly improving noise suppression capabilities in dark areas.

[0019] Third, existing Retinex methods force the RGB three channels to satisfy the constraint of "enhanced image × illumination = input image," which severely limits color recovery capabilities and easily leads to insufficient saturation or color cast. This invention innovatively proposes a grayscale space physical consistency constraint, maintaining the Retinex brightness degradation relationship only in the grayscale dimension, while introducing color cosine similarity loss to independently constrain hue consistency. While ensuring the physical rationality of brightness, this significantly improves the freedom of color recovery, effectively solving the problems of dull or color-cast colors in traditional methods, and enhancing the color naturalness of the results to a high degree of consistency with real-world scenes.

[0020] Fourth, the continuous Gaussian rendering architecture of this invention does not rely on a fixed pixel grid and can adapt to different resolution inputs. It achieves leading performance under various test conditions such as composite low-light scenes, real mobile phone shooting scenes, and professional SLR shooting scenes. It can stably output high-quality enhancement results in high-contrast backlight scenes, complex texture scenes, and high color fidelity demand scenes, and has strong generalization ability and practical application value.

[0021] 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

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 The present invention provides a step diagram of a low-light image enhancement method that integrates continuous Gaussian rendering and Retinex. Figure 2 This is a comparison image of the enhancement effect of the present invention on real-world backlight and low-light samples; Figure 3 This is a comparison chart showing the enhancement effect of complex textures and color card test samples in this invention. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] 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.

[0026] This invention discloses a low-light image enhancement method that integrates Continuous Gaussian Splatting (CGS) and Retinex, proposing an explicit-implicit joint modeling framework of CGS-Retinex. This is the first low-light enhancement scheme to deeply integrate Continuous Gaussian Splatting (CGS) with Retinex theory, fundamentally solving the bottlenecks of traditional methods such as discrete modeling artifacts, insufficient detail recovery, and color distortion. Figure 1 As shown, it includes the following steps: Step S1: Hierarchical Feature Encoding The input low-light image is subjected to hierarchical feature encoding to extract deep contextual features and shallow high-frequency skip features. The hierarchical feature encoder adopts a dense residual connection structure, and through multi-level downsampling and residual skip connections, it simultaneously extracts global contextual semantic features and local high-frequency detail features containing edge and texture information.

[0027] Given a low-light input image A hierarchical feature encoder with dense residual connections is introduced. Extracting deep contextual features and shallow skip features containing high-frequency spatial details :

[0028] Step S2: Generation of continuous Gaussian meta-parameter set Deep contextual features are input into a parallel illumination prediction head and a feature prediction head, which generate a continuous Gaussian parameter set for illumination and a continuous Gaussian parameter set for features, respectively. The illumination prediction head and the feature prediction head have a dual-branch parallel structure; the illumination prediction head is used to model the low-frequency continuous illumination parameter field, and the feature prediction head is used to model the high-frequency texture feature parameter field.

[0029] For feature tensors The prediction head first extracts basis features through parameter-shared nonlinear convolution, and then divides them into geometric and attribute branches. For each grid center point... The network needs to predict the corresponding Gaussian elements. .

[0030] First, in the normalized image coordinate system The base grid points for building internal rules The prediction network regresses a small offset in spatial location. Thus, the continuous center coordinates of each Gaussian element are obtained. :

[0031] Covariance matrix of Gaussian kernel This determines the range of influence and anisotropic shape of the scattering. To ensure the positive definiteness of the covariance matrix and facilitate the calculation of the Mahalanobis distance of the Gaussian distribution, the inverse of the covariance matrix is ​​directly predicted. This process is determined by the predicted anisotropy scaling factor. With rotation angle Composed of various elements. The scaling factor is achieved through... Activation function mapping is used to prevent singularities. Where:

[0032] Based on this, the rotation matrix and diagonal scaling matrix The product of these elements forms the transformation matrix, from which the specific elements of the inverse covariance matrix can be derived. :

[0033] The matrix elements can be precisely calculated using the following formula:

[0034] In addition, the network also predicts the opacity factor for each Gaussian unit. and characteristic attributes In the illumination prediction head In, attributes This represents the scalar illumination intensity of a single channel, and we use an adaptive activation function. Ensure its absolute nonnegativity. Meanwhile, in middle, Represents a high-dimensional semantic vector.

[0035] Step S3: Continuous Gaussian rendering By rendering the lighting continuous Gaussian meta-parameter set and the feature continuous Gaussian meta-parameter set separately using a continuous Gaussian renderer with learnable temperature, a spatially continuous global illumination field and context feature field are obtained.

[0036] The continuous Gaussian renderer extracts K×K neighboring Gaussian elements for each continuous query coordinate, calculates the Mahalanobis distance, and generates exponentially decaying weights to complete the weighted integration of the neighborhood Gaussian attributes. In this embodiment, K=3.

[0037] For a given continuous query coordinate The feature space is projected back and then expanded using the Unfold mechanism to extract its surrounding 3×3 neighborhood Gaussian set. .

[0038] against The Middle Gao Siyuan First, calculate the query point. Its center Spatial offset vector :

[0039] Furthermore, its squared distance to Mahalanobis is calculated. :

[0040] To prevent gradient explosion and control Gaussian tail, the distance term is truncated to its maximum value. Within this range, the exponentially decaying weight of the primitive is obtained:

[0041] The original mixing weights are calculated as the product of Gaussian decay and opacity:

[0042] The learnable temperature coefficient is The function is mapped to a positive adaptive temperature value to balance the smoothness and edge sharpness of the rendering result. To ensure the temperature term is strictly positive during optimization and to avoid division by zero errors, the temperature calculation uses a small offset. function:

[0043] Based on this adaptive temperature, the weights are adjusted within a local 3×3 neighborhood. Normalization operation:

[0044] Finally, the query point Attribute rendering value at location It is obtained by weighted integral of all neighborhood primitive attributes:

[0045] Among them, the illumination field The generation can be formalized as:

[0046] in, For continuous space Any query coordinate in the range.

[0047] This step also includes performing a boundary truncation operation on the global illumination field, constraining the minimum boundary of the illumination value to... To avoid numerical explosion and color bleaching caused by extremely low lighting values ​​during Retinex inverse operations:

[0048] Step S4: Calculation of the basic reflectivity map Based on the Retinex inverse operation, the fundamental reflectance map is calculated from the global illumination field and the input low-light image. Classical Retinex theory assumes that the observed image can be decomposed into a pixel-wise product of the illumination map and the reflectance map; therefore, the fundamental reflectance map is coarse. It can be directly derived through the classical Retinex inverse process:

[0049] Step S5: Reflectance Residual Map Prediction The context feature field, shallow high-frequency jump features, and spatial coordinate encoding are input into a coordinate-driven implicit neural network to predict reflectance residual maps. The coordinate-driven implicit neural network first maps two-dimensional spatial coordinates to a high-dimensional frequency space using a random Fourier basis to capture high-frequency texture information. Then, it fuses the global semantic information of the context feature field with the local detail information of the shallow high-frequency jump features. Finally, a multilayer perceptron outputs the reflectance residual values ​​at the corresponding coordinate positions, generating a complete reflectance residual map.

[0050] In order for a shallow MLP (Multilayer Perceptron) to sense high-frequency changes, a fixed random Fourier basis is first used. Normalized two-dimensional spatial coordinates Mapping to a higher-dimensional frequency space:

[0051] Subsequently, the high-frequency coordinates were encoded. The Gaussian renderer outputs spatially smooth and physically continuous contextual feature fields. And shallow skip features with resolution aligned via bilinear interpolation The concatenation is performed along the channel dimension. This combined tensor serves as the input to the implicit network:

[0052] To achieve joint nonlinear decoding of coordinates and high-dimensional features, a four-layer cascaded 1×1 convolutional network is used to reconstruct the texture residual. The mapping process of the first three hidden layers is abstracted as a composite nonlinear function. Feature tensor After hierarchical feature extraction, the last layer passes... Activation strictly constrains the residual output domain to [-1, 1]. The entire decoding map can be represented by the following condensed formula:

[0053] Among them, composite mapping This represents a nonlinear extraction process for deep features. Specifically, to ensure the statistical stability of spatial features, the initial hidden layer... It adopted group normalization and Activated convolution operation, i.e. The subsequent feature extraction layer and This removes normalization and adopts the standard. Convolution The joint operator.

[0054] Step S6: Final Enhanced Image Generation The base reflectance map and the reflectance residual map are linearly superimposed, and the final enhanced image is output after value range truncation. This step uses the additive residual Retinex paradigm to linearly fuse the base reflectance and reflectance residuals, and then... The function constrains the output to the [0,1] valid color space:

[0055] This innovative additive residual Retinex paradigm allows the INR network to focus on high-frequency texture compensation and color saturation correction in damaged dark areas without having to relearn the low-frequency color distribution of the image. This achieves an excellent balance between lighting smoothing and detail fidelity in principle.

[0056] Step S7: Network Training The present invention also includes a network training step: constructing a joint loss function based on multi-dimensional physical constraints to complete network training. The joint loss function includes Charbonnier reconstruction loss, SSIM loss, VGG perceptual loss, illumination total variation loss, grayscale physical consistency loss, and color cosine similarity loss.

[0057] Let the Ground Truth clear image be... The final output of the network prediction is The illumination diagram is as follows The calculation methods for each loss item are as follows: (1) Charbonnier reconstruction loss: In order to greatly improve the peak signal-to-noise ratio (PSNR) of the reconstructed image and encourage the network to focus on high-frequency details that are difficult to recover, Charbonnier Loss (smoothing) is adopted. Loss, as a fundamental constraint at the pixel level, is compared to Loss can effectively reduce edge blur:

[0058] in, This is a regularization minimum to avoid the gradient being singular at zero.

[0059] (2) SSIM Loss: To ensure the fidelity of spatial structure and local contrast, SSIM loss constraints are introduced:

[0060] (3) VGG Perceptual Loss: Deep feature representations are extracted using a pre-trained VGG-16 model, and mean squared error is calculated to align with the perceptual quality of the human visual system. To ensure consistency of feature dimensions, the image is first subjected to a preset mean. and variance Row instance normalization:

[0061] (4) Total variation loss of illumination: According to the physical priors of Retinex theory, the illumination distribution in nature should be piecewise smooth. This applies to illumination maps generated by continuous Gaussian predictors. Anisotropic total variation (TV) loss is applied to suppress potential local abrupt changes and high-frequency noise perturbations:

[0062] in, , This is the normalization coefficient for the spatial dimension.

[0063] (5) Color Cosine Similarity Loss: To combat the color shift phenomenon that is easily caused by enhancement under extremely dark conditions, a color consistency loss is introduced. Cosine similarity is calculated in the pixel channel dimension to narrow the angle between the predicted image and the label in the color vector space.

[0064] (6) Gray-scale physical consistency loss: A physical loss mechanism that only constrains consistency in the gray-scale space, while maintaining the Retinex degradation consistency of brightness, relaxes the restrictions on network color recovery. Let The table represents the mean pooling operation along the channel dimension:

[0065] Calculate the physical consistency loss using the Charbonnier loss function:

[0066] The overall optimization objective can be expressed as:

[0067] The hyperparameters are set as follows: .

[0068] Furthermore, to verify the effectiveness and generalization ability of the low-light image enhancement method proposed in this invention, this embodiment selects mainstream advanced methods in the field of low-light image enhancement as comparison benchmarks, and conducts comprehensive performance comparison experiments on publicly available synthetic datasets and real-world scene shooting datasets. The technical advantages of this invention are comprehensively verified by combining quantitative indicators and visualization effects.

[0069] To verify the effectiveness and generalization ability of the low-light image enhancement method proposed in this invention, this embodiment selects mainstream advanced methods in the field of low-light image enhancement as comparison benchmarks and conducts comprehensive performance comparison experiments on publicly available synthetic datasets and real-world scene shooting datasets. The technical advantages of this invention are comprehensively verified by combining quantitative indicators and visualization effects. This embodiment uses seven industry-standard evaluation metrics to quantitatively assess the enhancement effect. Three of these are full-reference metrics: Peak Signal-to-Noise Ratio (PSNR, dB; higher values ​​indicate better pixel-level fidelity), Structural Similarity (SSIM, range [0,1]; higher values ​​indicate better overall similarity in brightness, contrast, and structure), and Learned Perceptual Patch Similarity (LPIPS, range [0,1]; lower values ​​indicate less distortion at the human visual perception level). The four non-reference metrics are: Natural Image Quality Evaluator (NIQE; lower values ​​indicate a more realistic image), Luminance Order Error (LOE; lower values ​​indicate better preservation of the natural order of illumination before and after enhancement), Discrete Entropy (DE; higher values ​​indicate richer detail information in the image), and Enhancement Measurement Evaluation (EME; higher values ​​indicate better local contrast performance of the image).

[0070] Quantization comparison results on a real-world low-light mobile phone shooting dataset show that the method of this invention achieves an SSIM of 0.87, a PSNR of 20.19 dB, a low LOE of 4.30, a DE of 2.85, and an EME of 22.71. Among these, SSIM, LOE, DE, and EME rank first among all compared methods, and PSNR ranks second. Compared to the best-performing Retinex architecture method in the prior art, this invention improves SSIM by 40.3% and EME by 387.3%, demonstrating that this method possesses stronger color and detail recovery capabilities in complex low-light scenes captured by real mobile phones.

[0071] In complex scene tests on synthetic paired low-light datasets, the method of this invention achieved an SSIM of 0.92, a PSNR of 22.41 dB, a LPIPS as low as 0.11, and a NIQE as low as 3.86. All core evaluation metrics ranked first among all comparative methods. Compared with the best existing method, SSIM was improved by 10.8%, PSNR by 12.4%, and LPIPS was reduced by 26.7%, verifying the excellent reconstruction performance of this method in extremely complex low-light scenes.

[0072] Quantization comparison results on a real-world low-light mobile phone shooting dataset show that the method of this invention achieves an SSIM of 0.87, a PSNR of 20.19 dB, a low LOE of 4.30, a DE of 2.85, and an EME of 22.71. Among these, SSIM, LOE, DE, and EME rank first among all compared methods, and PSNR ranks second. Compared to the best-performing Retinex architecture method in the prior art, this invention improves SSIM by 40.3% and EME by 387.3%, demonstrating that this method possesses stronger color and detail recovery capabilities in complex low-light scenes captured by real mobile phones. In complex scene tests on synthetic paired low-light datasets, the method of this invention achieved an SSIM of 0.92, a PSNR of 22.41 dB, a LPIPS as low as 0.11, and a NIQE as low as 3.86. All core evaluation metrics ranked first among all comparative methods. Compared with the best existing method, SSIM was improved by 10.8%, PSNR by 12.4%, and LPIPS was reduced by 26.7%, verifying the excellent reconstruction performance of this method in extremely complex low-light scenes.

[0073] As attached Figure 2As shown, the comparison sample comes from a real-life backlit low-light scene by a window, including complex lighting conditions with a large contrast between bright buildings outside the window and dark railings inside. In terms of brightness performance, existing technologies such as EFI-NET, LIME, and RUAS generally suffer from insufficient brightness recovery, resulting in overall darkness of the buildings outside the window and difficulty in discerning details in the railing area. Methods such as NoiSER result in overexposure of the window area and loss of details in the sky. The brightness recovery of the method of this invention (CGS-Retinex) is highly consistent with the real reference image (Ground Truth). It can not only fully brighten dark areas such as railings, but also effectively suppress overexposure in the sky area. Details of building exteriors and vegetation are clearly visible, and the global illumination distribution is uniform and natural. In terms of color reproduction, methods such as EnlightenGAN and ZeroIG exhibit significant global color casts, with the overall image leaning towards green or blue. Methods like FMR-NET and PairLIE suffer from excessively high sky color saturation. The color reproduction of the method in this invention shows no significant color cast; the reddish-brown exterior walls of buildings, green vegetation, and the pale purple gradient of the sky are all essentially consistent with the real reference image, without color distortion or over-enhancement. Regarding sharpness, the enhancement results of methods such as ChebyLighter and SCI exhibit varying degrees of haze and texture blurring, with blurred building edge details and noticeable mesh artifacts in some methods. The enhancement results of the method in this invention are clear and bright, with sharp window frame edges, clearly discernible building wall textures and details of vegetation below, and no significant noise residue.

[0074] As attached Figure 3 As shown, the comparison sample comes from a complex texture and standard color chart test scene in a synthetic dataset, containing a large number of fine patterns and standard color test blocks: the comparison of the complex pattern area in the upper half shows that the enhancement results of existing techniques such as EFI-NET, SCI, and RUAS are either blurry with complete loss of color details of the pattern, or the noise is amplified and obvious graininess appears. Methods such as FRR-NET and NoiSER have obvious color casts and distorted pattern colors. The enhancement results of the method of this invention (CGS-Retinex) have clear pattern textures, sharp red, blue, and yellow color patterns, and color reproduction is highly consistent with the real reference image (GT), while there are no obvious noise artifacts. As can be seen from the comparison of the lower half of the standard color chart area, the enhancement results of existing technologies such as EFI-NET, FMR-NET, and RUAS are either severely distorted in color chart, with significant deviations in saturation and hue, or the details of dark color blocks are completely lost, presenting large areas of blurred color blocks. Some methods also amplify noise and result in severe graininess in the image. The method of this invention can accurately restore the hue and saturation of various standard colors such as pink, purple, blue, green, and yellow in the color chart, with sharp and clear edges of color blocks, while significantly suppressing noise in dark areas, without the jagged effect or color deviation commonly found in traditional methods.

[0075] Based on comprehensive quantitative indicators and visual performance, the low-light image enhancement method proposed in this invention, which integrates continuous Gaussian rendering and Retinex, surpasses existing mainstream technologies in all low-light scenes with high contrast ratios, complex textures, and high color fidelity requirements. It achieves the optimal balance between lighting smoothness, color fidelity, and detail recovery capabilities, and has strong generalization ability and practical application value.

[0076] 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 low-light image enhancement method integrating continuous Gaussian rendering and Retinex, characterized in that, Includes the following steps: S1: Perform hierarchical feature encoding on the input low-light image to extract deep contextual features and shallow high-frequency jump features; S2: Input the deep context features into the illumination prediction head and the feature prediction head to generate a continuous Gaussian parameter set for illumination and a continuous Gaussian parameter set for features; S3: By rendering the lighting continuous Gaussian meta-parameter set and the feature continuous Gaussian meta-parameter set separately using a continuous Gaussian renderer with learnable temperature, a spatially continuous global illumination field and context feature field are obtained. S4: Based on Retinex inverse operation, the basic reflectance map is calculated from the global illumination field and the input low-light image; S5: Input the context feature field, shallow high-frequency jump features and spatial coordinate encoding into the coordinates to drive the implicit neural network and predict the reflectance residual map; S6: Linearly superimpose the base reflectance map and the reflectance residual map, and output the final enhanced image after truncation of the value range.

2. The low-light image enhancement method according to claim 1, characterized in that, The hierarchical feature encoder described in step S1 adopts a dense residual connection structure, which simultaneously extracts global contextual semantic features and local high-frequency detail features containing edge and texture information through multi-level downsampling and residual skip connections.

3. The low-light image enhancement method of claim 1, wherein, The illumination prediction head and feature prediction head described in step S2 are a dual-branch parallel structure. The illumination prediction head is used to model the low-frequency continuous illumination parameter field, and the feature prediction head is used to model the high-frequency texture feature parameter field.

4. The low-light image enhancement method according to claim 1, characterized in that, The continuous Gaussian element parameter set mentioned in step S2 includes continuous center coordinates, inverse covariance matrix, opacity coefficient, and attribute parameters; wherein the inverse covariance matrix is ​​calculated by jointly using anisotropic scaling factor and rotation angle, and is used to characterize the spatial influence range and anisotropic shape of Gaussian elements.

5. The low-light image enhancement method of claim 1, wherein, The continuous Gaussian renderer described in step S3 extracts K×K neighborhood Gaussian elements for continuous query coordinates, calculates Mahalanobis distance and generates exponentially decaying weights, and completes the neighborhood Gaussian attribute weighted integral, where K is a natural number.

6. The low-light image enhancement method of claim 1, wherein, The learnable temperature coefficient mentioned in step S3 is mapped to a positive adaptive temperature value by the Softplus function, which is used to balance the smoothness and edge sharpness of the rendering result.

7. The low-light image enhancement method of claim 1, wherein, Step S3 also includes performing an adaptive truncation operation on the global illumination field, constraining the minimum boundary of the illumination value to 0.05, to avoid numerical explosion and color bleaching caused by division operations.

8. The low-light image enhancement method of claim 1, wherein, The coordinate-driven implicit neural network described in step S5 first maps two-dimensional spatial coordinates to a high-dimensional frequency space through a random Fourier basis to capture high-frequency texture information. Then, it fuses the global semantic information of the context feature field with the local detail information of shallow high-frequency jump features. The reflectance residual value corresponding to the coordinate position is output by the multilayer perceptron, and finally a complete reflectance residual map is generated.

9. The low-light image enhancement method of claim 1, wherein, Step S6 uses the additive residual Retinex paradigm to linearly fuse the base reflectance and the reflectance residual, and then uses the Clamp function to constrain the output within the [0,1] effective color space.

10. The low-light image enhancement method of claim 1, wherein, The method also includes step S7: Network training: The network training is completed by constructing a joint loss function based on multi-dimensional physical constraints. The joint loss function includes Charbonnier reconstruction loss, SSIM loss, VGG perceptual loss, illumination total variation loss, grayscale physical consistency loss and color cosine similarity loss, with the weights of each sub-item being 1.0, 0.5, 0.05, 0.05, 0.05 and 0.05, respectively.

Images