A low-light image enhancement method and device based on multi-feature collaborative modeling and a medium

Abstract

The present application relates to a kind of low-light image enhancement method, equipment and medium based on multi-feature collaborative modeling, comprising: after low-light image pre-processing, input dual-branch low-light image enhancement model, obtain enhanced HVI feature, convert HVI feature back to RGB color space, obtain enhanced image;Dual-branch low-light image enhancement model includes the dual-branch enhancement network of introducing illumination modulation module and double contrast enhancement module, the brightness branch and color structure branch of dual-branch enhancement network are updated in multiple scales cross, illumination modulation module self-adapting correction global illumination distribution, double contrast enhancement module generates enhancement chart and suppression chart;Dual-branch low-light image enhancement model executes color space conversion, dual-branch feature extraction, multiple scales cross update, illumination modulation and double contrast enhancement in turn.Compared with prior art, the present application can effectively improve low-light image brightness, while keeping color naturalness and structure details.

Description

Technical Field

[0001] This invention relates to the fields of image processing, deep learning and computer vision, and in particular to a low-light image enhancement method, device and medium based on multi-feature collaborative modeling. Background Technology

[0002] Low-light image enhancement is an important research area in computer vision. Its goal is to restore the brightness, texture details, and color information of an image under conditions of insufficient lighting, uneven exposure, or loss of detail in dark areas. Low-light images are widely present in scenarios such as nighttime shooting, indoor low-light conditions, surveillance videos, autonomous driving, and mobile photography.

[0003] In recent years, deep learning methods have become the mainstream technology for low-light enhancement. These methods learn the mapping relationship between low-light and normal-light images through end-to-end networks, achieving good results in brightness restoration and detail reconstruction. However, existing deep learning methods still have the following shortcomings: First, many methods directly enhance the image in the RGB color space, resulting in strong coupling between brightness and color information, which can easily lead to color shifts and enhancement artifacts. Second, most methods focus more on overall brightness enhancement and lack stable modeling of the global illumination distribution, leading to insufficient enhancement of dark areas or overexposure of bright areas in complex lighting environments. Third, existing methods lack sufficient modeling of local contrast and texture details, which may result in blurred details, unclear structure, or decreased visual naturalness in the enhanced image.

[0004] Chinese patent application CN119151808A discloses a training method and system for a low-light image enhancement model. The model comprises an enhancement network that enhances the brightness of low-light images and a calibration network that restores the brightness of low-light images. This model enhances the visual quality of low-light images. The model introduces a dynamic range fusion exposure suppression algorithm, which can effectively handle strong light noise that may occur at night and improve the visual effect of image enhancement. However, it is relatively coarse and cannot achieve stable modeling of the global illumination distribution. It still easily leads to insufficient enhancement of dark areas or overexposure of bright areas in complex lighting environments. Furthermore, it lacks sufficient modeling for local contrast and texture details, resulting in images that are still prone to blurring after enhancement.

[0005] In summary, how to effectively enhance dark areas, suppress bright areas, and restore local details while maintaining color naturalness and structural consistency remains an important problem to be solved in the field of low-light image enhancement. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a low-light image enhancement method, device and medium based on multi-feature collaborative modeling.

[0007] The objective of this invention can be achieved through the following technical solutions: A low-light image enhancement method based on multi-feature collaborative modeling, the method comprising: The low-light image to be enhanced is preprocessed and then input into a pre-trained dual-branch low-light image enhancement model to obtain the enhanced HVI features. The enhanced HVI features are then converted back to the RGB color space to obtain the enhanced image. The preprocessing includes image normalization and color space conversion, which converts the low-light image to be enhanced from the RGB space to the HVI space and decomposes it into a brightness branch and a color structure branch. The dual-branch low-light image enhancement model includes a dual-branch enhancement network that incorporates an illumination modulation module and a dual contrast enhancement module. The dual-branch enhancement network includes a luminance branch and a color structure branch. The luminance branch and the color structure branch achieve information interaction between luminance features and color structure features through a multi-scale cross-update mechanism. The illumination modulation module adaptively corrects the global illumination distribution, and the dual contrast enhancement module generates enhancement maps and suppression maps. The dual-branch low-light image enhancement model obtains enhanced HVI features by sequentially performing dual-branch feature extraction, multi-scale cross-updating, illumination modulation, and dual contrast enhancement.

[0008] Furthermore, the dual-branch enhancement network is constructed based on an encoder-decoder structure, and the luminance branch and color structure branch interact with each other at different scale levels of the dual-branch enhancement network; the scale levels include shallow, medium and deep scales. An illumination modulation module is embedded in the encoding layer of the luminance branch, and an illumination modulation module is embedded in the encoding layer of the color structure branch; a dual contrast enhancement module is embedded in the decoding layer of the luminance branch.

[0009] Furthermore, the dual contrast enhancement module is embedded in the middle layer of the luminance branch decoding layer scale level.

[0010] Furthermore, the process of color space conversion includes: The low-light image to be enhanced is subjected to pixel normalization processing, and the normalized low-light image is converted from RGB space to HVI space; The converted low-light image is decomposed into an I channel and an HV channel, where the I channel represents the image brightness information and the HV channel represents the image color and structure information. The I channel is input into the luminance branch, and the HV channel is input into the color structure branch, thus decoupling the luminance information from the color structure information.

[0011] Furthermore, the process of dual-branch feature extraction includes: Luminance features in the I channel are extracted and enhanced through the luminance branch; color and structural features in the HV channel are extracted and restored through the color structure branch. The multi-scale cross-update mechanism is set during the encoding and decoding process. During the multi-scale cross-update process, the brightness branch and color structure branch make the shallow features containing edge and texture information and the deep features containing semantic and global structural information complementary.

[0012] Furthermore, the illumination modulation is implemented based on an illumination modulation module, which includes a multi-layer convolution and a modulation head, wherein the modulation head is a two-layer convolution block with ReLU and Tanh activation; The process of light modulation includes: The feature map input from the illumination modulation module is received, and a normalization operation is performed to obtain the features. ; The feature The first branch and the second branch are processed in parallel. The features of the first branch are processed by a multi-head attention mechanism and a convolutional structure to model global dependencies and local contexts respectively, capturing the change patterns of lighting in the spatial dimension. The features of the second branch are processed by a feedforward network to enhance feature representation. The results from the two branches are added together to obtain the fusion feature. ; The fusion features The signal is fed into the modulation head to generate a single-channel modulation map. And Tanh constraints are used to ensure that the modulation constraints are within a controllable range; Based on the channel modulation mapping Residual modulation of the original features is performed to obtain the features after illumination enhancement.

[0013] Furthermore, the dual contrast enhancement is implemented based on a dual contrast enhancement module, which includes a multi-layer convolution, a decomposition module, and a fusion module; The dual contrast enhancement process includes: Receive shallow features from the encoder and same-scale deep features from the decoder Calculate the difference signal between the two; The differential signals are input into a convolutional network for information integration to obtain fused features. ; The fused features are fed into the decomposition module to obtain the enhancement map and the suppression map; The enhanced and suppressed images are fed into the fusion module to obtain the features after dual contrast enhancement.

[0014] Furthermore, the dual-branch low-light image enhancement model is pre-trained based on a comprehensive loss function; The expression for the comprehensive loss function is: in, To account for the overall losses, Main loss function, To assist in the loss weighting coefficient, As an auxiliary loss function; The expression for the main loss function is: in, These are weighted hyperparameters used to balance the losses in two different spaces. Used to measure the difference between two internal parameters. This is the HVI diagram generated after network enhancement. This is the HVI diagram obtained from the Ground Truth through spatial transformation. RGB image reconstructed from the network. RGB image of Ground Truth; The expression for the auxiliary loss function is: in, To enhance graph loss, To suppress graph loss, For the generated enhanced map, Input brightness, This is the generated suppression map.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the low-light image enhancement method based on multi-feature collaborative modeling as described above.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the low-light image enhancement method based on multi-feature collaborative modeling as described above.

[0017] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention effectively reduces color distortion and enhances image quality. Most existing low-light enhancement methods directly restore images in the RGB color space. Due to the high coupling between luminance and color information, color shift, local artifacts, and unnatural visual appearance are common problems during enhancement. This invention converts the input image to the HVI color space and decomposes it into luminance and color structure branches, enabling the network to model luminance and color structure information separately. This effectively reduces color distortion and luminance artifacts, improving the naturalness of the enhancement results.

[0018] 2. This invention achieves a balance between brightness restoration and structure preservation. It employs a dual-branch network structure to process brightness and color structure information separately, and introduces a cross-update mechanism at different scales, allowing brightness and structural features to complement each other. This design enhances overall brightness while preserving edges, textures, and object structures in the image, resulting in a clearer and more stable enhanced image.

[0019] 3. This invention enhances global illumination correction capabilities in complex lighting scenarios. Relying solely on implicit learning within a network is insufficient to achieve stable illumination restoration. This invention introduces an illumination modulation module, which adaptively adjusts the feature responses of different regions by generating spatial modulation maps. This enables the model to more accurately model the global illumination distribution, alleviating overall underexposure and uneven illumination.

[0020] 4. This invention can simultaneously enhance dark areas and suppress overexposure in bright areas. The invention proposes a dual contrast enhancement module that generates enhancement and suppression maps separately, performing differentiated processing on the dark areas to be enhanced and the bright areas to be suppressed in the image. This design can enhance dark textures and local contrast while suppressing excessive enhancement of bright areas, thereby improving the problem of coexisting underexposure and overexposure.

[0021] 5. The global illumination modulation and local contrast enhancement of this invention have complementary effects, resulting in a more stable and natural enhancement effect. The illumination modulation module in this invention is mainly used for global illumination distribution correction, while the dual contrast enhancement module is mainly used for local contrast and detail optimization. Both improve low-light degradation problems from both global and local perspectives, forming a complementary and synergistic relationship. Through this synergistic mechanism, this invention can simultaneously handle multiple degradation phenomena such as overall underexposure, uneven illumination, loss of detail in dark areas, and overexposure in highlights, making the enhanced image more stable in terms of brightness, color, structure, and visual naturalness. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a diagram of the enhanced network structure of the method of the present invention.

[0024] Figure 3 This is a structural diagram of the illumination modulation module of the method of the present invention.

[0025] Figure 4 This is a structural diagram of the dual contrast enhancement module of the method of the present invention.

[0026] Figure 5 The image shows a comparison of the results of the low-light image enhancement method designed for multi-feature collaborative modeling in this invention with the results of several existing methods. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] Example 1 This embodiment discloses a low-light image enhancement method based on multi-feature collaborative modeling, the method as follows: Figure 1 As shown, steps S1-S3 are included, and each step is described in detail below: Step S1: Preprocess the low-light image to be enhanced. The preprocessing includes image normalization and color space conversion.

[0029] The input image is first normalized and then converted from the RGB color space to the HVI color space. This conversion yields an I channel and an HV channel, where the I channel primarily represents the image's luminance information, and the HV channel primarily represents the image's color and structural information. This decomposition method reduces the interference between luminance and color information during low-light enhancement, allowing the network to model luminance recovery and color structure preservation separately, thus reducing color shift, artifacts, and unnatural enhancement phenomena.

[0030] Step S2: Input the preprocessed image into the dual-branch low-light image enhancement model to obtain the enhanced HVI features.

[0031] The dual-branch low-light image enhancement model obtains enhanced HVI features by sequentially performing dual-branch feature extraction, multi-scale cross-updating, illumination modulation, and dual contrast enhancement.

[0032] The dual-branch low-light image enhancement model includes a dual-branch enhancement network that incorporates an illumination modulation module and a dual contrast enhancement module.

[0033] The dual-branch augmentation network includes a luminance branch and a color structure branch. The luminance branch and the color structure branch realize the information interaction between luminance features and color structure features through a multi-scale cross-update mechanism.

[0034] The dual-branch augmentation network is built on an encoder-decoder structure. The luminance branch and the color structure branch exchange information at different scale levels in the dual-branch augmentation network, including shallow, medium and deep scales.

[0035] Specifically, such as Figure 2 As shown, the luminance branch is used to extract and enhance the luminance features in the I channel, while the color structure branch is used to extract and recover the color and structural features in the HV channel. At different scale levels of the network, the two branches exchange information through a multi-scale cross-update mechanism, enabling luminance features to provide illumination guidance for color structure recovery, while color structure features provide edge and texture constraints for luminance enhancement.

[0036] The dual-branch feature extraction process includes: extracting and enhancing the luminance features in the I channel through the luminance branch; and extracting and restoring the color and structural features in the HV channel through the color and structure branch.

[0037] An illumination modulation module is embedded in the coding layer of the luminance branch; an illumination modulation module is embedded in the coding layer of the color structure branch; and a dual contrast enhancement module is embedded in the decoding layer of the luminance branch. Specifically, the dual contrast enhancement module is embedded in the middle layer of the luminance branch decoding layer scale level.

[0038] A multi-scale cross-update mechanism is implemented during encoding and decoding to pass complementary information across shallow, mid-level, and deep features. Shallow features contain richer edge and texture information, while deep features contain stronger semantic and global structural information. Through multi-scale cross-update, the two branches can maintain semantic consistency during enhancement, avoiding the destruction of color and structural information by enhancing brightness alone.

[0039] An illumination modulation module is introduced into the augmentation network to adaptively correct the global illumination distribution: Illumination modulation is implemented based on an illumination modulation module, which includes multi-layer convolution and a modulation head. The modulation head is a two-layer convolution block with ReLU and Tanh activation.

[0040] In the augmented network, to avoid premature corruption by high-level semantics or extensive downsampling, two illumination modulation modules are embedded in the coding layer of the luminance branch. Simultaneously, because low-light degradation not only affects luminance but also indirectly reduces the preservation of image color and structure through luminance distortion, an illumination modulation module is embedded in the coding layer of the color and structure branch. This provides illumination-consistent feature inputs for subsequent decoding stages in both branches.

[0041] The purpose of the illumination modulation module is to adaptively adjust the illumination distribution in the feature space. Instead of directly estimating the explicit illumination map, it learns a spatial modulation map and implicitly adjusts the brightness of each spatial region, thereby achieving stable and visually perceptual global illumination correction.

[0042] like Figure 3 As shown, the process of light modulation includes: For the input feature map First, a normalization operation is performed to obtain... These features are then processed in parallel by two branches. The first branch models global dependencies and local contexts using a multi-head attention mechanism and a convolutional structure, respectively, to capture the spatial variation patterns of illumination. The second branch enhances the feature representation using a feedforward network. Finally, the two branches are added together to obtain the fused features. : in, express, express.

[0043] Subsequently It is fed into the modulation head, which is a two-layer convolutional block with ReLU and Tanh activations, capable of generating a single-channel modulation map. The Tanh constraint ensures that the modulation constraint is within a controllable range, preventing excessive brightness amplification. .

[0044] Based on channel modulation mapping Residual modulation of the original features is performed to obtain the output features after illumination enhancement. : in, It is a learnable scalar, initialized to 0.1, used to control the modulation intensity, and applied to the original features after element-wise multiplication.

[0045] The illumination modulation module features implicit global illumination correction, avoiding artifacts caused by direct decomposition; it ensures numerical stability through residual modulation; and it has both global and local adaptive capabilities, effectively handling complex non-uniform illumination distributions.

[0046] A dual contrast enhancement module is introduced, which generates an enhancement map and a suppression map respectively, to enhance details in dark areas and suppress overexposure in bright areas. Dual contrast enhancement is implemented based on a dual contrast enhancement module, which includes multi-layer convolution, decomposition, and fusion modules.

[0047] In the enhancement network, the dual contrast enhancement module operates by learning the difference in contrast, which is reflected in changes in brightness. Its purpose is to adjust the magnitude and direction of brightness enhancement. Combining these two factors, it is embedded in the decoding layer of the brightness branch. Furthermore, within the decoding layer, it is embedded in the intermediate stage because, in the intermediate stage, dark and bright areas are still separable, and edges and structures have not yet been over-compressed.

[0048] The dual contrast enhancement module achieves adaptive optimization of local image contrast by explicitly modeling enhancement and suppression regions. Its core idea is to analyze the differences between shallow and deep features in the network and generate two complementary spatial modulation maps based on this: the enhancement map enhances the detail in underexposed areas, while the suppression map attenuates the effects of overly bright or noisy areas. This dual modulation mechanism enables the network to finely adjust the contrast distribution of each region in a spatially aware manner.

[0049] The process of dual contrast enhancement is as follows Figure 4 As shown, it includes: The dual contrast enhancement module takes two features as input: shallow features from the encoder. and same-scale deep features from the decoder .

[0050] First, calculate the difference signal between the two. This difference reflects the changes in lighting and texture understanding between early and later layers of the network, highlighting areas that need further enhancement or suppression. The differential signal is then input into a small convolutional network for information integration to obtain fused features. : .

[0051] The fused features are then fed into the decomposition module to obtain enhancement and suppression maps. Finally, the enhancement map, suppression map, and deep features are fed into the fusion module to obtain the final features. : in, and It is a learnable parameter used to control the strength of enhancement and suppression. The initial value is 0.1, and it is automatically updated through backpropagation.

[0052] By separating enhancement and suppression behaviors, the dual contrast enhancement module effectively constrains bright area responses while enhancing dark details. Unlike traditional pixel-level adjustments, the dual contrast enhancement module operates in a deep semantic feature space, enabling it to achieve semantically perceptive local contrast correction based on image content. The module employs a residual structure to ensure training stability and consistency of feature representation. This module complements the illumination modulation module, which is responsible for correcting global illumination distribution, while the dual contrast enhancement module focuses on refined enhancement of local contrast and texture structure. During model training, the generated enhancement and suppression maps also serve as supervisory signals in loss calculation, guiding the network to learn local contrast adjustment strategies more accurately.

[0053] The dual-branch low-light image enhancement model is pre-trained based on a comprehensive loss function.

[0054] During model training, low-light images are used as input, and the overall loss function includes the main loss function. and auxiliary loss function The main loss function constrains the difference between the enhanced image and the normally illuminated image output by the network. The main loss function includes HVI color space loss and RGB color space loss to simultaneously constrain the brightness, color, and structure restoration effects of the image. It is a weighted hyperparameter.

[0055] The expression for the comprehensive loss function is: in, To account for the overall losses, Main loss function, To assist in the loss weighting coefficient, This is an auxiliary loss function.

[0056] The expression for the main loss function is: in, These are weighted hyperparameters used to balance the losses in two different spaces. Used to measure the difference between two internal parameters. This is the HVI diagram generated after network enhancement. This is the HVI diagram obtained from the Ground Truth through spatial transformation. RGB image reconstructed from the network. RGB image of Ground Truth; The auxiliary loss, serving as a key guiding signal for training the dual contrast enhancement module, aims to constrain the generated enhancement map to have a stronger response in dark areas and the suppression map to have a stronger response in bright areas. Its expression is as follows: The input brightness is normalized to [0,1].

[0057] The expression for the auxiliary loss function is: in, To enhance graph loss, To suppress graph loss, For the generated enhanced map, Input brightness, This is the generated suppression map.

[0058] After training, the low-light image to be enhanced is input into the trained model. First, the input image is converted to the HVI color space, and brightness features and color structure features are extracted respectively. Then, feature enhancement is performed through a dual-branch network, a multi-scale cross-update mechanism, an illumination modulation module, and a dual contrast enhancement module. Finally, the enhanced HVI features are converted back to the RGB color space to obtain an enhanced image with improved brightness, natural colors, and clear structural details.

[0059] Step S3: Convert the enhanced HVI features back to the RGB color space to obtain the enhanced image.

[0060] The final output is an enhanced image with improved brightness, natural colors, and clear structural details.

[0061] Example 2 This embodiment, based on Embodiment 1 above, discloses an experimental verification of a low-light image enhancement method based on multi-feature collaborative modeling, used to verify the superiority of the technical solution in Embodiment 1, and comparative experiments are conducted on three datasets: LOL, LoLI-Street, and RE2L.

[0062] The LOL dataset is a widely used benchmark dataset in the field of low-light image enhancement. Its core feature is that it provides strictly aligned pairs of low-light and normal-light images in real-world scenes, providing a crucial training foundation for supervised learning methods and serving as an important standard for validating low-light enhancement models.

[0063] LoLI-Street contains a large number of image pairs from modern urban street scenes, covering 33,000 pairs of data under low-light and normal-light conditions, which are suitable for robustness verification of models in complex real-world scenarios.

[0064] RE2L is a benchmark dataset for low-light image enhancement focusing on real-world outdoor scenes. It captures the changes in light at different times and exposure levels through natural lighting, providing richer lighting information as guidance for model training.

[0065] All experiments were conducted on an NVIDIA RTX A5000 GPU platform, with the model implemented using the PyTorch framework. During training, the Adam optimizer was used, with cropSize set to 256×256, batchSize to 8, and an initial learning rate of 1e-4. Learning rate scheduling employed a periodic cosine annealing strategy with a warmup phase: first, a warmup phase, followed by periodic reduction of the learning rate to a minimum value according to a cosine function, and then restoration to a higher learning rate, forming a cyclic decay of training cycles. The model was trained for a total of 600 epochs, and the final model was selected based on the best performance on the validation set. For evaluation metrics, PSNR and SSIM were used as the basic indicators for measuring image reconstruction quality. Furthermore, to further evaluate the quality of the augmentation results at the perceptual level, the LPIPS metric was introduced, which better reflects the perceptual characteristics of human visual systems to image differences. Several existing algorithms were compared, and the results are shown in Table 1.

[0066] Table 1 Comparison data of multiple algorithms This embodiment integrates global illumination modulation and local contrast optimization mechanisms. In the HVI color space, the two mechanisms work together to enhance detail fidelity while maintaining brightness consistency, thereby jointly improving problems such as underexposure in dark areas, overexposure in highlights, and insufficient texture recovery, making the enhancement result more visually natural and structurally complete.

[0067] Experimental results show that this method maintains stable enhancement performance and high visual quality under different lighting conditions. Tests were conducted on three datasets, such as... Figure 5 As shown, the image generated in this embodiment outperforms the other three comparison methods in terms of overall lighting and color restoration, as well as preservation of local details. This indicates that the technical solution of this embodiment improves low-light enhancement performance and produces a more natural and detailed image.

[0068] Example 3 Based on Embodiment 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the low-light image enhancement method based on multi-feature collaborative modeling as described above.

[0069] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the aforementioned low-light image enhancement method based on multi-feature collaborative modeling. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0070] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0071] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0072] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A low-light image enhancement method based on multi-feature collaborative modeling, characterized in that, The method includes: The low-light image to be enhanced is preprocessed and then input into a pre-trained dual-branch low-light image enhancement model to obtain the enhanced HVI features. The enhanced HVI features are then converted back to the RGB color space to obtain the enhanced image. The preprocessing includes image normalization and color space conversion, which converts the low-light image to be enhanced from the RGB space to the HVI space and decomposes it into a brightness branch and a color structure branch. The dual-branch low-light image enhancement model includes a dual-branch enhancement network that incorporates an illumination modulation module and a dual contrast enhancement module. The dual-branch enhancement network includes a luminance branch and a color structure branch. The luminance branch and the color structure branch achieve information interaction between luminance features and color structure features through a multi-scale cross-update mechanism. The illumination modulation module adaptively corrects the global illumination distribution, and the dual contrast enhancement module generates enhancement maps and suppression maps. The dual-branch low-light image enhancement model obtains enhanced HVI features by sequentially performing dual-branch feature extraction, multi-scale cross-updating, illumination modulation, and dual contrast enhancement.

2. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 1, characterized in that, The dual-branch enhancement network is constructed based on an encoder-decoder structure, and the luminance branch and color structure branch interact with each other at different scale levels of the dual-branch enhancement network. The scale hierarchy includes shallow, intermediate, and deep scales; An illumination modulation module is embedded in the coding layer of the luminance branch, and an illumination modulation module is embedded in the coding layer of the color structure branch. A dual contrast enhancement module is embedded in the decoding layer of the luminance branch.

3. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 2, characterized in that, The dual contrast enhancement module is embedded in the middle layer of the luminance branch decoding layer scale level.

4. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 1, characterized in that, The process of color space conversion includes: The low-light image to be enhanced is subjected to pixel normalization processing, and the normalized low-light image is converted from RGB space to HVI space; The converted low-light image is decomposed into an I channel and an HV channel, where the I channel represents the image brightness information and the HV channel represents the image color and structure information. The I channel is input into the luminance branch, and the HV channel is input into the color structure branch, thus decoupling the luminance information from the color structure information.

5. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 4, characterized in that, The process of dual-branch feature extraction includes: Luminance features in the I channel are extracted and enhanced through the luminance branch; color and structural features in the HV channel are extracted and restored through the color structure branch. The multi-scale cross-update mechanism is set during the encoding and decoding process. During the multi-scale cross-update process, the brightness branch and color structure branch make the shallow features containing edge and texture information and the deep features containing semantic and global structural information complementary.

6. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 1, characterized in that, The illumination modulation is implemented based on an illumination modulation module, which includes a multi-layer convolution and a modulation head. The modulation head is a two-layer convolution block with ReLU and Tanh activation. The process of light modulation includes: The feature map input from the illumination modulation module is received, and a normalization operation is performed to obtain the features. ; The feature The first branch and the second branch are processed in parallel. The features of the first branch are processed by a multi-head attention mechanism and a convolutional structure to model global dependencies and local contexts respectively, capturing the change patterns of lighting in the spatial dimension. The features of the second branch are processed by a feedforward network to enhance feature representation. The results from the two branches are added together to obtain the fusion feature. ; The fusion features The signal is fed into the modulation head to generate a single-channel modulation map. And Tanh constraints are used to ensure that the modulation constraints are within a controllable range; Based on the channel modulation mapping Residual modulation of the original features is performed to obtain the features after illumination enhancement.

7. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 1, characterized in that, The dual contrast enhancement is implemented based on a dual contrast enhancement module, which includes a multi-layer convolution, a decomposition module, and a fusion module. The dual contrast enhancement process includes: Receive shallow features from the encoder and same-scale deep features from the decoder Calculate the difference signal between the two; The differential signals are input into a convolutional network for information integration to obtain fused features. ; The fused features are fed into the decomposition module to obtain the enhancement map and the suppression map; The enhanced and suppressed images are fed into the fusion module to obtain the features after dual contrast enhancement.

8. The low-light image enhancement method based on multi-feature collaborative modeling according to claim 1, characterized in that, The dual-branch low-light image enhancement model is pre-trained based on a comprehensive loss function; The expression for the comprehensive loss function is: in, To account for the overall losses, Main loss function To assist in the loss weighting coefficient, As an auxiliary loss function; The expression for the main loss function is: in, These are weighted hyperparameters used to balance the losses in two different spaces. Used to measure the difference between two internal parameters. This is the HVI diagram generated after network enhancement. This is the HVI diagram obtained from the Ground Truth through spatial transformation. RGB image reconstructed from the network. RGB image of Ground Truth; The expression for the auxiliary loss function is: in, To enhance graph loss, To suppress graph loss, For the generated enhanced map, Input brightness, This is the generated suppression map.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the low-light image enhancement method based on multi-feature collaborative modeling as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the low-light image enhancement method based on multi-feature collaborative modeling as described in any one of claims 1-8.

Images