Low-illumination image enhancement and deblurring combined restoration method

Abstract

The invention discloses a low-illumination image enhancement and deblurring combined restoration method in the technical field of computer vision and image processing, and the method comprises the steps: decoupling brightness and color information in an HVI color space, and constructing a double-branch neural network; a dynamic feature perception enhancement module is introduced into an HV branch, and a non-uniform fuzzy region in an image is adaptively processed by generating a position-related dynamic convolution kernel; constructing a structure perception enhancement sub-network, explicitly extracting high-frequency components and edge features of the image, and realizing self-adaptive fusion of multi-source features through a structure guide fusion block so as to reconstruct clear textures and contours; through joint optimization of consistency loss, high-frequency loss and edge loss of HVI and RGB spaces, brightness, color and structure information of an image are recovered cooperatively. According to the method, effective decoupling of the brightness and the color is realized, the non-uniform fuzzy processing capability is improved, the recovery details of the structure and the texture are enhanced, and the robustness is high.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a method for combined low-light image enhancement and deblurring restoration. Background Technology

[0002] In low-light environments (such as nighttime surveillance, nighttime driving, and handheld night scene photography), acquired images typically face a dual degradation challenge: low contrast and high noise due to insufficient light, and motion blur caused by long exposures or camera shake. This combined degradation severely impacts the accuracy of subsequent advanced vision tasks (such as object detection and semantic segmentation).

[0003] In existing technologies, there are two main types of solutions for dealing with this type of problem:

[0004] The first type is the cascaded approach, which connects existing low-light enhancement algorithms and deblurring algorithms sequentially. However, this "step-by-step" strategy ignores the coupling between tasks. If enhancement is performed before deblurring, the enhancement process often amplifies noise and introduces overexposure, making it difficult to estimate the blur kernel; if deblurring is performed before enhancement, the blur features under low light are difficult to extract accurately, and the artifacts introduced by deblurring will be amplified by subsequent enhancement steps.

[0005] The second type is a joint processing scheme based on the RGB color space, which attempts to perform enhancement and deblurring simultaneously in a single network. However, the three channels of the RGB color space are highly correlated (coupled), making it easy to disrupt the color balance when adjusting brightness, resulting in color casts or color distortion in the enhanced image.

[0006] Furthermore, existing end-to-end network models typically tend to fit the low-frequency components of an image, while lacking the ability to recover severely damaged high-frequency textures and edge structures, resulting in blurred details and unclear outlines in the restored image.

[0007] Based on this, the present invention designs a method for combined low-light image enhancement and deblurring restoration to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for joint restoration of low-light image enhancement and deblurring. By decoupling brightness and color in the HVI space and introducing a dynamic feature perception and structure guidance mechanism, high-quality restoration of low-light blurred images is achieved, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for joint image enhancement and deblurring restoration in low-light conditions includes the following steps:

[0011] S1. Obtain the low-light blurred RGB image to be processed, convert it to HVI color space, and decouple the image into HV components containing color and structural information and I components containing brightness information.

[0012] S2. Construct a two-branch neural network model containing an I branch and an HV branch. The I branch is used for brightness estimation and correction, and the HV branch is used for the recovery of color and structural information. The two branches interact with each other through a cross-attention module.

[0013] S3. In the HV branch, a position-dependent dynamic convolution kernel is generated using a dynamic feature perception enhancement module. This module uses encoder features to predict spatially changing convolution kernels, performs pointwise convolution on decoder features, and performs spatial adaptive enhancement on feature maps to handle non-uniform blur.

[0014] S4. Construct a structure-aware enhancement subnetwork between the encoder and decoder of the HV branch. This subnetwork contains a high-frequency modeling branch and an edge modeling branch, which extract the high-frequency features and edge features of the image respectively, and reconstruct the damaged structure and texture in the low-light blurred image.

[0015] S5. Using a structure-guided fusion block, a coordinate attention-based mechanism is adopted to adaptively fuse the high-frequency features and edge features extracted in step S4 into the backbone features of the network, guiding the decoder to recover image details.

[0016] S6. Perform inverse HVI transformation on the enhanced HV component and I component of the network output to obtain the restored RGB image, and train and optimize the network by calculating the joint loss function including color consistency loss, high frequency loss and edge loss.

[0017] Furthermore, the image decoupling in step S1 includes:

[0018] By utilizing the characteristics of the HVI color space, brightness information, which is easily affected by lighting, is separated into the I channel for separate processing, while information containing image texture and color attributes is retained in the HV plane for enhancement, so as to avoid color distortion during the enhancement process.

[0019] Furthermore, the dual-branch network in step S2 has a lightweight cross-attention module between the I branch and the HV branch, which is used to perform interaction of brightness and color features at different resolution levels.

[0020] Furthermore, the generation of position-dependent dynamic convolutional kernels using the dynamic feature-aware enhancement module in step S3 includes:

[0021] S3-1. Predicting dynamic convolutional kernels using multi-scale enhanced features from encoder output. ;

[0022] S3-2, For each spatial location on the feature map Generate a size of dynamic kernel ;

[0023] S3-3. In the decoder, the features to be enhanced are extracted using a sliding window method. In position Neighborhood characteristics;

[0024] S3-4. Apply the generated dynamic convolution kernel to perform pointwise convolution on the neighborhood features to obtain the enhanced features. The calculation formula is as follows:

[0025]

[0026] in, For the number of channels, For the size of the dynamic kernel, This represents the index coordinates within the convolution kernel and the feature block. This indicates a summation operation.

[0027] This dynamic feature perception enhancement module generates adaptive representations through cross-layer interaction, thereby more precisely repairing spatial non-uniform fuzziness.

[0028] Furthermore, the construction of the structure-aware enhancement subnetwork in step S4 includes:

[0029] S4-1. Deep HV features extracted by the encoder As input, deep representation modeling is performed using residual blocks to obtain... ;

[0030] S4-2, will The high-frequency modeling branch and the edge modeling branch are input separately. Each of these branches contains several residual blocks, and they predict the high-frequency features of the structure-aware enhancement output. With edge features ;

[0031] S4-3. During the training phase, the predicted high-frequency and edge features are subjected to inverse HVI transform to convert them back to the RGB color space for establishing regularization constraints. The real reference image is downsampled by 8 times in the RGB space to extract the real high-frequency and edge maps, which serve as the regression supervision targets for the high-frequency and edge modeling branches. This allows the network to operate at low resolution, thereby reducing computational overhead while maintaining reconstruction capabilities.

[0032] Furthermore, the structure-guided fusion block in step S5 employs a coordinate attention-based mechanism and includes:

[0033] S5-1, Determine the deep main features Respectively with high frequency characteristics and edge features Add them together to get and The summed features are then subjected to average pooling along both the horizontal and vertical directions to obtain the receptive vectors in both directions. and :

[0034]

[0035]

[0036] in, This is the perception vector in the vertical direction obtained after average pooling along the horizontal direction. This is the perception vector in the horizontal direction obtained after average pooling along the vertical direction;

[0037] S5-2. Concatenate the receptive vectors from the two directions in the spatial dimension, and generate an intermediate feature map through convolutional layers, normalization layers, and the ReLU activation function. :

[0038]

[0039] in, This is a reduction factor used to decrease the channel dimension;

[0040] S5-3. Segment the intermediate feature maps along the spatial direction and generate high-directional attention maps by applying convolution and sigmoid activation respectively. and wide-direction attention map ;

[0041] S5-4. Use attention maps to weight the features, and then add and fuse the weighted high-frequency feature branches and edge feature branches with the main features to output the final structure-enhanced features.

[0042] Furthermore, the formula for generating the attention map and weighted features in step S5 is as follows:

[0043]

[0044]

[0045]

[0046] in, and The intermediate feature tensor after segmentation. This represents the Sigmoid activation function. This indicates element-wise multiplication. The high-frequency features are enhanced; the edge feature branches are processed in the same way as the high-frequency feature branches.

[0047] Furthermore, the joint loss function in step S6 Defined as:

[0048]

[0049] in, For the loss of consistency in the RGB color space, For HVI color space consistency loss, For high-frequency information loss, This results in the loss of edge information. and For weight hyperparameters.

[0050] Furthermore, the high-frequency information loss in step S6 With edge information loss The calculation method is as follows:

[0051]

[0052]

[0053] in, and These are the high-frequency map and edge map predicted by the structure-aware enhancement subnetwork and output through inverse HVIT transform, respectively. This is a high-frequency image extracted from a real image using discrete cosine transform; This is an edge map extracted from a real image using the Canny operator; This represents the L1 distance.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. This invention achieves effective decoupling of brightness and color: by adopting the HVI color space and designing a dual-branch architecture, the interference of brightness adjustment on color information in the RGB space is avoided, significantly reducing color distortion in the enhancement process, and making the restored image color more natural and accurate.

[0056] 2. This invention improves the processing capability of non-uniform blur: It introduces a dynamic feature perception enhancement module, which breaks the limitation of traditional static convolution kernels and can generate adaptive dynamic kernels according to the blur differences of different regions of the image (such as moving objects and static backgrounds), effectively removing non-uniform motion blur.

[0057] 3. This invention enhances the details of structure and texture restoration: High-frequency and edge features are explicitly extracted and supervised through a structure-aware enhancement subnetwork, and structure-guided fusion blocks are used to integrate them into the backbone network, which solves the problems of texture loss and unclear contours under low light, resulting in a restored image with rich details and sharp edges.

[0058] 4. Strong robustness: The method of this invention can simultaneously handle the combined degradation of insufficient lighting and motion blur, and has good generalization ability in complex scenes. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a diagram of the overall network architecture of the method of the present invention;

[0061] Figure 2 This is a schematic diagram of the dynamic feature perception enhancement module in the method of the present invention;

[0062] Figure 3 This is a schematic diagram of the structure-guided fusion block in the method of the present invention;

[0063] Figure 4 This illustrates the effect of applying the method of the present invention to images containing clothing and fabrics. Detailed Implementation

[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please refer to the accompanying drawings. This invention provides a technical solution:

[0066] A joint restoration method for low-light image enhancement and deblurring can effectively decouple brightness and color, and can explicitly utilize structural information to guide the restoration.

[0067] I. Methods and Steps

[0068] Step 1: Data Preprocessing and Spatial Transformation

[0069] The low-light blurred RGB image to be processed is obtained, and the RGB image is converted to HVI space using the characteristics of HVI color space. HVI space consists of hue, value and intensity.

[0070] In this embodiment, the input image is first normalized to the [0,1] interval. The RGB image is then decomposed into two parts using the HVI transformation formula: the I component corresponds to the brightness information of the image, with a data dimension of H×W×1; the HV component corresponds to the color and structure information of the image, with a data dimension of H×W×2.

[0071] This decoupling allows subsequent networks to specifically address insufficient lighting (for the I component) and structural blurring and color distortion (for the HV component).

[0072] Step 2: Construct a two-branch network model

[0073] like Figure 1 As shown, this invention constructs a dual-branch network based on the U-Net architecture, as detailed below:

[0074] The I branch is mainly responsible for brightness estimation and correction, and the encoder part includes four stages:

[0075] First stage: Input I component (H×W×1) is expanded to 36 channels through a 3×3 convolutional layer;

[0076] The second to fourth stages: Downsampling is performed through convolution and bilinear sampling operations, followed by PReLU activation function and layer normalization. The feature map size becomes H / 2×W / 2, H / 4×W / 4 and H / 8×W / 8 respectively, and the number of channels becomes 36, 72 and 144 respectively.

[0077] The decoder consists of four stages. Upsampling is performed using convolutional layers and bilinear interpolation, gradually restoring the feature map size. At each resolution level, the decoder features are skipped to the corresponding features of the encoder, using channel concatenation.

[0078] The HV branch is responsible for the recovery of color and structural information. Its basic architecture is similar to that of the I branch, but special modules are introduced at key nodes. The input is a complete HVI three-channel component of H×W×3. The encoder and decoder also contain a four-level hierarchical structure, and the number of channels in each level is consistent with that of the I branch (36, 36, 72, 144).

[0079] A lightweight cross-attention module is introduced after each downsampling in the encoder and before each upsampling in the decoder in both branches. In the I branch, the lightweight cross-attention module receives features from the I branch as query conditions and features from the HV branch as keys and values; in the HV branch, it receives features from the HV branch as query conditions and features from the I branch as keys and values. Through an attention mechanism, it calculates the weighted mutual influence, thereby enabling the interaction of luminance and color features at different resolution levels.

[0080] Step 3: Non-uniform blurring processing (dynamic feature perception enhancement)

[0081] In the HV branch, dynamic feature perception enhancement is implemented to address non-uniform motion blur commonly seen in low-light conditions. The specific implementation process is as follows: Figure 2 As shown:

[0082] Step 3.1: Utilize the feature E (size E) output by the encoder. This method predicts the coefficient map of a dynamic convolutional kernel through a series of convolutional layers. The number of channels in this coefficient map is... , where d is the size of the dynamic kernel (d=5 in this embodiment). The output tensor size is .

[0083] Step 3.2: Reshape the above coefficient graph as follows The six-dimensional tensor corresponds to each location in space. Each has an independent convolution kernel .

[0084] Step 3.3: For the features to be enhanced in the decoder (size is) ), using a sliding window to extract the surrounding area of ​​each pixel. Features of the neighborhood The corresponding size is .

[0085] Step 3.4: Generate the dynamic kernel With extracted neighborhood features Perform element-wise multiplication and summation operations. The formulas are as follows:

[0086]

[0087] In this way, the network can generate a specific convolution kernel for each pixel based on the degree and direction of blur in the local area of ​​the image, thereby effectively removing non-uniform blur.

[0088] Step 4: Explicit extraction and reconstruction of structural features

[0089] At the encoder output of the HV branch (deep features) ,size Access structure-aware enhanced subnetwork.

[0090] Step 4.1: First, it is processed through 5 cascaded residual blocks to obtain... Each residual block consists of "convolution-ReLU activation-convolution" and contains residual connections.

[0091] Step 4.2: Two-branch modeling. High-frequency modeling branch input. High-frequency features are extracted using three residual blocks. The edge modeling branch also inputs... Edge features are extracted using three residual blocks. .

[0092] Step 4.3: During the training phase, the predicted high-frequency features and edge features are subjected to inverse HVI transform to convert them back to the RGB color space for establishing regularization constraints. The real reference image is downsampled by a factor of 8 in the RGB space. Discrete cosine transform is then performed on the downsampled image to extract the real high-frequency map, and the Canny operator is used to extract the real edge map. These two real images are used as labels, and the results are calculated... and The L1 loss forces the network to learn structural information.

[0093] Step 5: Integrate the extracted structural features into the backbone network using a structure-guided fusion block.

[0094] The specific implementation process is as follows: Figure 3 As shown:

[0095] Step 5.1: Feature Fusion Preparation. Prepare deep master features. Respectively with high frequency characteristics and edge features By adding element by element, we get and .

[0096] Step 5.2: Coordinate attention calculation (taking high-frequency branches as an example), for Average pooling is performed along the horizontal direction to obtain a size of eigenvectors; for Average pooling is performed along the vertical direction to obtain a size of The feature vectors are obtained by concatenating the vectors from the two directions in the spatial dimension, compressing the number of channels to C / r (r=8 in this example) through a 1×1 convolutional layer, and then performing batch normalization and ReLU activation to obtain the intermediate features. The intermediate features are re-segmented into horizontal and vertical parts, and each part is subjected to a 1×1 convolution to restore the number of channels to C. Finally, a sigmoid function is applied to activate the feature, generating a high-directional attention map. and wide-direction attention map .

[0097] Step 5.3: Weight the original features using the generated attention map:

[0098]

[0099] Step 5.4: Weight the high-frequency features Weighted edge features Compared with original deep features The features are added together to output the final structural enhancement features, which are then input into the subsequent decoder.

[0100] Step 6: Image Generation and Joint Optimization

[0101] The network's HV branch outputs the enhanced HV component, and the I branch outputs the corrected I component. These two components are then combined and synthesized into the final RGB image using the inverse HVI transform formula.

[0102] During training, a joint loss function is used. Perform end-to-end network optimization:

[0103]

[0104] in: and The L1 reconstruction loss is calculated for RGB and HVI spaces respectively, ensuring global consistency of color and brightness. To account for high-frequency information loss, the L1 distance between the predicted high-frequency features and the true high-frequency map is calculated. To account for edge information loss, the L1 distance between the predicted edge features and the ground truth edge map is calculated. In this embodiment, the weight parameters are set to... =1, =0.01.

[0105] In summary, this invention creatively combines HVI color space decoupling and structure-aware enhancement strategies to solve the long-standing problems of brightness-color coupling and detail loss in low-light image enhancement and deblurring tasks.

[0106] When given a low-light blurred image as input, the system first converts it to the HVI space. Through the I branch, the network uses a lightweight cross-attention module to accurately estimate the illumination distribution at different scales, solving the overexposure problem caused by brightness enhancement in traditional methods.

[0107] Meanwhile, the HV branch focuses on color and detail restoration, in which the dynamic feature perception enhancement module plays a crucial role. It overcomes the limitation of traditional static convolution kernels in dealing with spatially varying blur by generating position-dependent dynamic kernels, accurately repairing motion blur and non-uniform blur in images.

[0108] More importantly, this invention designs a structure-aware enhancement subnetwork to address the high-frequency information that is easily lost under low-light conditions. Unlike existing technologies that rely solely on implicit learning, the structure-aware enhancement subnetwork forces the network to "pay attention" to those textures and contours that are blurred in the dark through explicit high-frequency and edge supervision.

[0109] The structure-guided fusion block uses a coordinate attention mechanism to intelligently fuse these valuable structural priors back into the backbone network, ensuring that the decoder can perform pixel-level repair based on edge guidance when reconstructing images.

[0110] Ultimately, under the constraints of the joint loss function, this invention not only achieves a natural increase in brightness, but also reshapes clear and sharp image details while maintaining color fidelity, significantly improving the image quality in both subjective visual and objective metrics, and providing a high-quality data foundation for subsequent advanced vision tasks.

[0111] II. Simulation Experiment and Experimental Results

[0112] 1. Experimental Environment and Setup

[0113] This invention evaluates the effectiveness and performance of the proposed method through experiments on synthetic and real-world datasets. The experimental hardware platform was a workstation equipped with an NVIDIA RTX 4090 GPU. The software environment was based on the PyTorch deep learning framework. During training, the Adam optimizer was used with parameters set to β1=0.9 and β2=0.999. The initial learning rate was set to 1e−4 and gradually decayed to 1e−7 during training using a cosine annealing strategy.

[0114] For input image preprocessing, the image is randomly cropped into 256×256 pixel blocks. The training batch size is set to 4, and the total number of training epochs is 400. Regarding the setting of the loss function weights and hyperparameters... and Set them to 1 and 0.01 respectively.

[0115] The experiment used synthetic datasets (LOL-Blur, LOL-v1, LOL-v2) and real-world datasets (Real-LOL-Blur, DICM, LIME, MEF, NPE, VV) for comprehensive evaluation.

[0116] The objective evaluation indicators used include:

[0117] (1) Full reference metrics (for datasets with reference images): Peak signal-to-noise ratio (PSNR, in dB), structural similarity index (SSIM), and learned perceptual patch similarity (LPIPS). Among them, higher PSNR and SSIM are better, and lower LPIPS are better.

[0118] (2) No-reference metrics (for real no-reference datasets): MUSIQ, NRQM, NIQE, and BRISQUE. Among them, higher scores for MUSIQ and NRQM indicate better quality, while lower scores for NIQE and BRISQUE indicate better quality.

[0119] 2. Experimental Results and Analysis

[0120] (1) Combined low-light enhancement and deblurring performance

[0121] To verify the performance of this invention in handling composite degradation (low illumination + blur) tasks, the method of this invention was compared with existing cascaded methods (such as Zero-DCE cascaded MIMO) and joint processing methods (such as LEDNet and CIDNet) on the LOL-Blur dataset. The experimental results are shown in Table 1:

[0122] Table 1

[0123]

[0124] As shown in Table 1, the method of this invention outperforms the comparative methods in both SSIM (0.902) and LPIPS (0.093). This indicates that the present invention, through HVI spatial decoupling and structure-aware strategies, has significant technical advantages in maintaining image structural integrity and improving perceptual quality.

[0125] In addition, to verify the generalization ability of the method, tests were conducted on the real dataset Real-LOL-Blur, and the results are shown in Table 2:

[0126] Table 2

[0127]

[0128] The results in Table 2 show that the present invention performs well in NRQM and NIQE metrics, indicating that the present invention can still effectively restore image sharpness and color contrast in real and complex scenarios.

[0129] (2) Low-light image enhancement performance

[0130] To verify the universality of the present invention in a single low-light enhancement task, tests were conducted on synthetic datasets (LOL-v1, LOL-v2) and real-world scene datasets (DICM, LIME, etc.), and the comparison results are shown in Tables 3, 4 and 5, respectively.

[0131] Table 3

[0132]

[0133] Table 4

[0134]

[0135] Table 5

[0136]

[0137] The above data shows that the method of the present invention can maintain a stable enhancement effect on both synthetic and real data, especially its advantages in SSIM and NIQE indicators, which verifies the effectiveness of the present invention in maintaining the naturalness and structural information of images.

[0138] (3) Validation of key modules

[0139] To verify the specific technical effects of the key technical features (dynamic feature perception enhancement module, feature fusion strategy, and structural information modeling) in this invention, the following intra-group comparative experiments were designed.

[0140] To verify the contribution of the dynamic feature perception enhancement module to the deblurring performance, a comparative experiment with different configurations was set up, and the results are shown in Table 6.

[0141] Table 6

[0142]

[0143] Experimental data shows that the model achieves the best performance balance when the dynamic feature perception enhancement module is applied to the HV branch responsible for color and structure. This verifies that the technical approach of using dynamic kernels to process non-uniform blur in HV space is correct.

[0144] To verify that the structure-guided fusion block fusion strategy proposed in this invention is superior to traditional fusion methods, a comparison was conducted as shown in Table 7. In the first method, we concatenate the image, high-frequency features, and edge features along the channel dimension, and then use a 3×3 convolutional layer to perform feature fusion and output the result. This method is named "Convolutional Direct Fusion Method". Based on this, we added a coordinate attention module to simultaneously consider channel information and positional information, achieving adaptive feature fusion. This scheme is called "Attention-Based Direct Fusion Method". In addition, we constructed two cascaded fusion baselines: one is the "High-Frequency-Then-Edge Fusion Method", which first combines the image with high-frequency features, and then performs a secondary fusion with edge features; the other is the "Edge-Then-High-Frequency Fusion Method", which first combines the image with edge features, and then performs a secondary fusion with high-frequency features.

[0145] Table 7

[0146]

[0147] Experiments revealed that the "direct fusion method with attention" significantly outperforms the "convolutional direct fusion method," while the other two cascaded fusion methods are ineffective. In contrast, our proposed structure-guided fusion block demonstrates a clear advantage across all metrics. Data proves that the coordinate attention-based structure-guided fusion method employed in this invention, compared to direct stitching or cascaded fusion, can more effectively utilize high-frequency and edge information, thereby significantly improving image restoration quality.

[0148] To verify the necessity of explicitly introducing high-frequency and edge supervision signals in the structure-aware enhancement subnetwork, tests were conducted on the use of high-frequency information alone, edge information alone, and their combined use. The results are shown in Table 8.

[0149] Table 8

[0150]

[0151] Experimental results show that both high-frequency and edge information contribute positively to image restoration, and the combined use of both yields the best results, improving the peak signal-to-noise ratio by 0.53 dB compared to the benchmark. This fully demonstrates that explicit modeling of structural features in HVI space can effectively compensate for the loss of detail under low illumination.

[0152] (4) The application effect of the method of the present invention on clothing and fabric images

[0153] To further verify the practicality and generalization ability of the method of the present invention in complex scenarios, a restoration experiment was conducted specifically on low-light blurred images containing clothing, fabrics, and home textiles. Under low-light indoor conditions at night, clothing images are prone to problems such as dull colors, blurred textures, and loss of pattern details, which directly affect visual perception and subsequent analysis (such as e-commerce display, fabric quality inspection, virtual try-on, and other applications).

[0154] We selected several low-light blurred images containing elements such as clothing, wallpaper, doll fabric, and embroidery fabric from the dataset as test samples. These images generally suffer from combined degradation due to insufficient illumination, motion blur, and noise interference. After processing with the method of this invention, a significant quality improvement was achieved, such as... Figure 4 As shown.

[0155] The restoration solution provided by this invention can significantly improve the quality of clothing and fabric images captured in low-light environments, making their colors more realistic, details richer, and outlines clearer. This not only improves the visual experience for users browsing product images on e-commerce platforms and social media, but also provides higher-quality input data for advanced visual tasks such as image-based clothing design assistance, automatic fabric defect detection, and accurate modeling for virtual fitting systems, thus having significant practical application value.

[0156] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0157] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for joint restoration of low-light image enhancement and deblurring, characterized in that, Includes the following steps: S1. Obtain the low-light blurred RGB image to be processed, convert it to HVI color space, and decouple the image into HV components containing color and structural information and I components containing brightness information. S2. Construct a two-branch neural network model containing an I branch and an HV branch, where the I branch is used for brightness estimation and correction, and the HV branch is used for the recovery of color and structural information. S3. In the HV branch, the dynamic feature perception enhancement module is used to generate position-dependent dynamic convolution kernels to spatially adaptively enhance the feature map in order to handle non-uniform blur. S4. Construct a structure-aware enhancement subnetwork between the encoder and decoder in the HV branch to extract high-frequency features and edge features of the image, and reconstruct damaged structures and textures in low-light blurred images. S5. Using the structure-guided fusion block, the high-frequency features and edge features extracted in step S4 are adaptively fused into the backbone features of the network to guide the decoder to recover image details. S6. Perform inverse HVI transformation on the enhanced HV component and I component of the network output to obtain the restored RGB image, and train and optimize the network by calculating the joint loss function.

2. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The image decoupling in step S1 includes: By utilizing the characteristics of the HVI color space, brightness information, which is easily affected by lighting, is separated into the I channel for separate processing, while information containing image texture and color attributes is retained in the HV plane for enhancement, so as to avoid color distortion during the enhancement process.

3. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The dual-branch network in step S2 has a lightweight cross-attention module between the I branch and the HV branch, which is used to perform interaction of brightness and color features at different resolution levels.

4. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The step S3, which utilizes the dynamic feature perception enhancement module to generate position-dependent dynamic convolutional kernels, includes: S3-1. Predicting dynamic convolutional kernels using multi-scale enhanced features from encoder output. ; S3-2, For each spatial location on the feature map Generate a size of dynamic kernel ; S3-3. In the decoder, the features to be enhanced are extracted using a sliding window method. In position Neighborhood features ; S3-4. Apply the generated dynamic convolution kernel to perform pointwise convolution on the neighborhood features to obtain the enhanced features. The calculation formula is as follows: in, For the number of channels, For the size of the dynamic kernel, This represents the index coordinates within the convolution kernel and the feature block. This indicates a summation operation.

5. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The construction of the structure-aware enhancement subnetwork in step S4 includes: S4-1. Deep HV features extracted by the encoder As input, deep representation modeling is performed using residual blocks to obtain... ; S4-2, will The high-frequency modeling branch and the edge modeling branch are input separately. Each of these branches contains several residual blocks, and they predict the high-frequency features of the structure-aware enhancement output. With edge features ; S4-3. During the training phase, the predicted high-frequency features and edge features are subjected to inverse HVI transformation and converted back to the RGB color space to establish regularization constraints. The real reference image is downsampled by 8 times in the RGB space to extract the real high-frequency map and the real edge map, which are used as the regression supervision targets for the high-frequency modeling branch and the edge modeling branch.

6. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The structure-guided fusion block in step S5 employs a coordinate attention-based mechanism and includes: S5-1, Determine the deep main features Respectively with high frequency characteristics and edge features Add them together to get and The summed features are then subjected to average pooling along both the horizontal and vertical directions to obtain the receptive vectors in both directions. and : in, This is the perception vector in the vertical direction obtained after average pooling along the horizontal direction. This is the perception vector in the horizontal direction obtained after average pooling along the vertical direction; S5-2. Concatenate the receptive vectors from the two directions in the spatial dimension, and generate an intermediate feature map through convolutional layers, normalization layers, and the ReLU activation function. : in, This is a reduction factor used to decrease the channel dimension; S5-3. Segment the intermediate feature maps along the spatial direction and generate high-directional attention maps by applying convolution and sigmoid activation respectively. and wide-direction attention map ; S5-4. Use attention maps to weight the features, and then add and fuse the weighted high-frequency feature branches and edge feature branches with the main features to output the final structure-enhanced features.

7. The low-light image enhancement and deblurring joint restoration method according to claim 6, characterized in that, The formula for generating the attention map and weighted features in step S5 is as follows: in, and The intermediate feature tensor after segmentation. This represents the Sigmoid activation function. This indicates element-wise multiplication. The high-frequency features are enhanced; the edge feature branches are processed in the same way as the high-frequency feature branches.

8. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, The joint loss function in step S6 Defined as: in, For the loss of consistency in the RGB color space, For HVI color space consistency loss, For high-frequency information loss, This results in the loss of edge information. and For weight hyperparameters.

9. The low-light image enhancement and deblurring joint restoration method according to claim 1, characterized in that, High-frequency information loss in step S6 With edge information loss The calculation method is as follows: in, and These are the high-frequency map and edge map predicted by the structure-aware enhancement subnetwork and output through inverse HVIT transform, respectively. This is a high-frequency image extracted from a real image using discrete cosine transform; This is an edge map extracted from a real image using the Canny operator; This represents the L1 distance.

Images