A complex environment image restoration method based on degradation perception and multi-scale adaptive enhancement

Abstract

This invention proposes a method for restoring complex environment images based on degradation awareness and multi-scale adaptive enhancement. First, a degradation-aware style feature learning mechanism is designed: a dual-branch structure is used to simultaneously output style feature vectors and degradation degree predictions. A continuous contrastive learning objective is constructed based on structural similarity to obtain fine-grained style representations. Second, a multi-scale adaptive feature enhancement mechanism is proposed: at high-resolution scales, details and structures are enhanced separately through high- and low-frequency decomposition; at low-resolution scales, multi-dilation rate convolution and spatial pyramid self-attention are fused to capture local and global context. Experiments show that on the Outdoor-Rain, RainDrop, and Snow100K-L datasets, the proposed method achieves PSNR / SSIM of 30.24dB / 0.9111, 31.89dB / 0.9265, and 30.70dB / 0.9060, respectively, with an average PSNR of 30.94dB and an average SSIM of 0.9145. This invention significantly improves the restoration quality of complex weather-degraded images by employing scale-based enhancement and fine-grained style conditions.

Description

Technical Field

[0001] This method belongs to the field of computer vision and image processing technology, specifically involving a complex environment image restoration method based on degradation perception and multi-scale adaptive enhancement. Addressing the core problem of severe image quality degradation caused by multiple degradation factors such as severe weather (rain, raindrops, snow) and motion blur, existing methods mostly focus on single degradation types, lacking unified modeling capabilities for multiple degradation factors. Furthermore, style feature extraction is limited to global statistical information, and multi-scale feature processing is simplistic and difficult to adapt to real-world complex scenes. The overall approach of this method is as follows: First, a degradation-aware style feature learning mechanism is used to extract fine-grained style representations that integrate degradation type and degree. Then, a multi-scale adaptive feature enhancement mechanism is employed to enhance high-frequency details and low-frequency structures according to resolution. Specifically, this includes a dual-branch style network combined with structural similarity contrastive learning, and collaborative processing of high- and low-frequency decomposition and multi-dilation rate pyramid self-attention. This method can significantly improve the restoration quality of complex weather-degraded images and is suitable for practical applications such as smart cities and autonomous driving. Background Technology

[0002] This method belongs to the field of computer vision, specifically the sub-topic of image restoration. Image restoration has wide applications in common scenarios such as smart city security monitoring, autonomous driving, aerial remote sensing, medical image analysis, and industrial inspection. However, due to limitations such as imaging equipment performance and environmental constraints, the acquired image data often suffers from various quality problems, including rain lines, raindrops, snow spots, haze, low light, motion blur, and sensor noise. Therefore, the core objective of image restoration is to reconstruct high-quality, high-fidelity, clear images from these degraded observation images. From a technical implementation perspective, existing methods in this field are mainly divided into two categories: the first category is traditional methods based on physical prior modeling, which includes three subcategories: filtering-based, optimization-based, and physical model-based; the second category is data-driven deep learning methods, which include four subcategories: convolutional neural networks, generative adversarial networks, Transformer architecture, and diffusion models. These subcategories will be introduced in the following text in order.

[0003] 1. Traditional methods based on physical prior modeling

[0004] Traditional methods based on physical prior modeling mainly include filtering-based, optimization-based, and physics-based methods. The core idea of ​​these methods is to rely on signal processing theory and manually designed statistical priors or physical models to restore clear content from degraded images through mathematical transformations, without relying on large-scale training data.

[0005] (1) Filtering-based methods

[0006] The core principle of this type of method is to use statistical information within the pixel neighborhood to smooth or enhance the image, and to suppress degenerative components by replacing the central pixel value with the pixel value calculation within the local window.

[0007] Among them, median filtering is a classic nonlinear filtering method, and its basic form is shown in equation (1):

[0008]

[0009] in, The input is a degraded image. In pixels The neighborhood window centered on the center (a certain parameter). This method involves taking the median. It effectively suppresses salt-and-pepper noise while preserving edge information better than linear filtering. Variations of this method include weighted median filtering and recursive median filtering. The former assigns different weights to pixels at different locations to improve edge preservation, while the latter improves denoising through multiple iterations.

[0010] Another common method is bilateral filtering, which combines Gaussian weights in the spatial domain and the pixel value domain to smooth noise while preserving edge structure. Its definition is shown in equation (2):

[0011]

[0012] in, and Gaussian kernel functions for the spatial domain and range, respectively. and For the corresponding standard deviation parameter, The normalization factor is used. This method can preserve image edges well while denoising, but it is often ineffective for non-Gaussian and non-uniform weather degradation (such as rain lines, haze, and snow spots), and is prone to over-smoothing of textures. To further improve performance, researchers proposed guided filtering, which uses guided images to calculate linear transformation coefficients, and can effectively smooth textures while preserving edges (Reference 1: He, Kaiming, Jian Sun, and XiaoouTang. 2013. “Guided Image Filtering.” *IEEE Transactions on Pattern Analysis and Machine Intelligence* 35 (6): 1397–1409. doi:10.1109 / TPAMI.2012.213.). The contribution of this type of method is its high computational efficiency and simple principle; its limitation is that it relies heavily on manually designed filter kernels and is difficult to adaptively model complex and varied degradation patterns. Therefore, it is often combined with another type of data-driven method to improve performance.

[0013] (2) Optimization model-based methods

[0014] This type of method transforms the image restoration problem into an optimization problem by constructing an energy function that includes a data fidelity term and a priori regularization term. A clear image is obtained by solving this optimization problem. Depending on the regularization term, it can be divided into several subtypes.

[0015] The total variational model is one of the most representative methods. Its core idea is to assume that the sharp image is sparsity in the gradient domain, thereby causing the restoration result to exhibit piecewise smoothness. Its energy functional is shown in equation (3):

[0016]

[0017] The first term is the total variation regularization term (which promotes piecewise smoothing), the second term is the fidelity term (which constrains the closeness of the restored result to the observed data), and the parameters are... To observe the image, For the image to be recovered, To balance a certain constant (Reference 2: Rudin, Leonid I., Stanley Osher, and Emad Fatemi. 1992. “Nonlinear TotalVariation Based Noise Removal Algorithms.” *Physica D: Nonlinear Phenomena*60 (1-4): 259–268. doi:10.1016 / 0167-2789(92)90242-F.), this model can effectively suppress Gaussian noise and preserve image edges, but it often produces a staircase effect or excessively smoothed textures for non-Gaussian, non-uniform weather degradation (such as rain lines, haze, and snow spots). To alleviate this problem, researchers proposed a higher-order total variational model, which introduces second-order derivative information to reduce the staircase effect.

[0018] Sparse representation is another important class of methods. It assumes that image patches can be linearly combined with a small number of non-zero coefficients on an overcomplete dictionary, and the image is recovered by solving a norm optimization problem (Reference 3: Wright, John, Allen Y. Yang, Arvind Ganesh, S. Shankar Sastry, and Yi Ma. 2009. “Robust Face Recognition via Sparse Representation.” *IEEE Transactions on Pattern Analysis and Machine Intelligence* 31 (2): 210–227. doi:10.1109 / TPAMI.2008.79.). Its optimization objective is shown in Equation (4):

[0019]

[0020] in, To extract the first A matrix of image patches, For the image to be recovered, For an overcomplete dictionary (a certain parameter). For sparse coding coefficients, This is the regularization parameter (a constant). Because... Norm is difficult to solve, so it is usually relaxed to Norms. Variations of this method include K-SVD dictionary learning and nonlocal sparse representation, the latter leveraging the self-similarity of images to further improve reconstruction quality. The contribution of this class of methods lies in their good theoretical interpretability and lack of reliance on large amounts of training data; their limitations include limited ability to model structural redundancy under complex weather patterns, and the computationally intensive optimization process.

[0021] (3) Physical model-based methods

[0022] These methods establish physical imaging models for specific degradation processes and then use the estimation of key parameters in the model to inversely solve for a clear image. Different degradation types correspond to different physical models.

[0023] In image dehazing tasks, the atmospheric scattering model is the most classic physical model, and its core form is shown in Equation (5):

[0024]

[0025] in, For the observed foggy images, For the clear image to be recovered, For global atmospheric light (a certain parameter). Transmittance (a certain parameter). Atmospheric scattering coefficient, The scene depth is given by the dark channel prior method. Based on this model, the dark channel prior method found through statistical analysis of a large number of haze-free images that in most non-sky local areas, at least one color channel pixel value approaches zero. Based on this, the transmittance and atmospheric light can be estimated (Reference 4: He, Kaiming, Jian Sun, and Xiaoou Tang. 2011. “Single Image Haze Removal Using Dark Channel Prior.” *IEEE Transactions on Pattern Analysis and Machine Intelligence* 33 (12): 2341–2353. doi:10.1109 / TPAMI.2010.168.). The transmittance estimation formula is shown in Equation (6):

[0026]

[0027] in, This is a constant (usually 0.95) used to retain a small amount of fog to maintain the natural look of the image. For The method uses a local window centered on the color attenuation prior and a nonlocal prior to address dehazing requirements in different scenarios. Subsequent improvements to this method include color attenuation priors and nonlocal priors.

[0028] In image deraining tasks, physical models typically represent rain maps as a linear superposition of the background and rain streaks, as shown in Equation (7):

[0029]

[0030] in, For the observed rain map, Background image, The rain streak component is represented by a dictionary learning or low-rank decomposition method. Based on this model, researchers have separated the two components using dictionary learning or low-rank decomposition (Reference 5: Kang, Li-Wei, Chia-Wen Lin, and Yu-Hsiang Fu. 2012. “Automatic Single-Image-Based Rain Streaks Removal via Image Decomposition.” *IEEE Transactions on Image Processing* 21 (4): 1742–1755. doi:10.1109 / TIP.2011.2179057.). The contribution of this type of method lies in its clear physical meaning and good performance for single degradation types (such as haze and rain streaks); its limitation is that it is highly sensitive to the accuracy of the physical model. When the actual degradation process does not match the model assumptions (such as rain and snow mixing, haze and raindrop coexisting scenarios), the recovery quality will drop sharply.

[0031] 2. Data-driven deep learning methods

[0032] Data-driven deep learning methods mainly include those based on convolutional neural networks, generative adversarial networks, Transformer architectures, and diffusion models. These methods are characterized by learning end-to-end mappings from degraded to sharp images from large-scale datasets by constructing deep neural networks, gradually replacing manually designed prior models.

[0033] (1) Convolutional Neural Network-Based Methods

[0034] Convolutional Neural Networks (CNNs) efficiently extract image features through local connectivity and weight sharing mechanisms, becoming the foundational architecture for image restoration tasks. The definitions of conventional convolution operations are shown in equations (8a) and (8b):

[0035]

[0036]

[0037] in, For the input feature map, For convolution kernel, This outputs a feature map. Dilated convolution expands the receptive field by inserting holes between kernel elements; its effective kernel size is... ,in For the inflation rate (see reference 6: Yu, Fisher, and Vladlen Koltun. 2016. “Multi-Scale Context Aggregation by Dilated Convolutions.” *Proceedings of the International Conference on Learning Representations (ICLR)*.).

[0038] A typical model of this type of method is SRCNN (Super-Resolution Convolutional Neural Network), which was the first to apply deep networks to image super-resolution tasks, proving the effectiveness of end-to-end learning. Its core structure consists of three convolutional layers, which are responsible for image patch extraction and representation, nonlinear mapping, and final reconstruction, respectively (Reference 7: Dong, Chao, Chen Change Loy, Kaiming He, and XiaoouTang. 2014. “Learning a Deep Convolutional Network for Image Super-Resolution.” *Proceedings of the European Conference on Computer Vision (ECCV)*, 184–199. doi:10.1007 / 978-3-319-10593-2_13.).

[0039] Based on this, the encoder-decoder structure is widely adopted. It extracts contextual features with a larger receptive field through progressive downsampling, and then restores spatial resolution through progressive upsampling to generate a detailed output image (Reference 8: Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. 2015. “U-Net: Convolutional Networks for Biomedical Image Segmentation.” *Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)*, 234–241. doi:10.1007 / 978-3-319-24574-4_28.). Its core operation can be expressed as Equation (10):

[0040]

[0041] in, and These represent downsampling (stretch convolution or pooling) and upsampling (transposed convolution or bilinear interpolation) operations, respectively. This represents feature fusion (usually channel concatenation or element-wise addition). The design of skip connections allows the decoder to utilize both shallow detail information and deep semantic information simultaneously.

[0042] Residual networks alleviate the training difficulties of deep networks by introducing skip connections, allowing for a significant increase in network depth (Reference 9: He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. “Deep Residual Learning for Image Recognition.” *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)*, 770–778. doi:10.1109 / CVPR.2016.90.). The basic form of its residual block is shown in Equation (10):

[0043]

[0044] in, This is the output feature map of the residual block. For input, The residual mapping to be learned (typically consisting of two convolutional layers and an activation function) is used. This architecture has been shown to effectively improve reconstruction quality in image restoration tasks and allows for the construction of deep networks with more than 100 layers.

[0045] Densely connected networks further extend the idea of ​​residual connections by directly connecting each layer to all previous layers, promoting feature reuse and gradient flow (Reference 10: Huang, Gao, Zhuang Liu, Laurens Van DerMaaten, and Kilian Q. Weinberger. 2017. “Densely Connected Convolutional Networks.” *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)*, 4700–4708. doi:10.1109 / CVPR.2017.243.). In image restoration, this type of network contributes to strong feature learning capabilities, achieving significant success in single degradation tasks (such as denoising, deblurring, and deraining); its limitation lies in its insufficient generalization ability for complex and varied multi-weather degradation scenarios.

[0046] (2) Generative Adversarial Network-based methods

[0047] This type of method generates visually more realistic reconstructions through adversarial training between the generator and the discriminator (Reference 11: Isola, Phillip, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. 2017.“Image-to-Image Translation with Conditional Adversarial Networks.” *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)*, 1125–1134. doi:10.1109 / CVPR.2017.632.). Its basic idea originates from generative adversarial networks, which introduce adversarial losses to encourage the generator to produce realistic images that are difficult for the discriminator to distinguish.

[0048] The overall optimization objective of the generative adversarial network is shown in Equation (11):

[0049]

[0050] in, For the generator (responsible for generating clear images in the recovery task). It is a discriminator (responsible for distinguishing between real, clear images and generated images). For real samples, The input is a noisy or degraded image. In conditional generative adversarial networks, the generator additionally receives conditional information. (e.g., degraded image), the objective function becomes equation (12):

[0051]

[0052] In image restoration tasks, the generator typically employs an encoder-decoder structure or a residual network structure, while the discriminator is responsible for determining the authenticity of the restored result. To improve restoration quality, pixel-level losses (such as...) are often incorporated. or The methods employ both perceptual loss (based on feature matching loss from a pre-trained VGG network) and feature matching loss (based on feature matching loss from a pre-trained VGG network). These methods can generate visually more natural and texture-rich restored images, overcoming the drawback of traditional pixel-level loss functions that tend to result in overly smooth images. However, their limitations include unstable training processes, susceptibility to pattern collapse, and the potential introduction of unrealistic artifacts in the generated results.

[0053] (3) Transformer-based approach

[0054] The Transformer architecture initially achieved great success in the field of natural language processing, its core being the self-attention mechanism (reference 12: Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” *Advances in Neural Information Processing Systems (NeurIPS)* 30.). The self-attention mechanism is defined as shown in equation (13):

[0055]

[0056] in, , , These are the query matrix, key matrix, and value matrix, respectively. Given the input feature sequence, Let be the dimension of the key (a certain parameter). Multi-head self-attention executes the above process in parallel on multiple independent representation subspaces to capture different types of dependencies, and its output is shown in Equation (14):

[0057]

[0058] in, The number of attention heads (a certain parameter). , To output the fusion matrix.

[0059] The Transformer encoder layer is followed by a feedforward network after the self-attention sublayer, and is equipped with residual connections and layer normalization. Its forward propagation process is shown in equations (15) and (16):

[0060]

[0061]

[0062] in, It is a feedforward network.

[0063] In the field of computer vision, ViT (Vision Transformer) was the first to divide an image into fixed-size patches and input these patches as a sequence into a Transformer encoder (Reference 13: Dosovitskiy, Alexey, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, et al. 2021. “An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale.” *Proceedings of the International Conference on Learning Representations (ICLR)*.). Subsequently, the Swin Transformer expanded its receptive field while maintaining computational efficiency by introducing a shifted window self-attention mechanism (Reference 14: Liu, Ze, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. “Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows.” *Proceedings of the IEEE / CVF International Conference on Computer Vision (ICCV)*, 10012–10022. doi:10.1109 / ICCV48922.2021.00986.).

[0064] In the field of image restoration, IPT (Image Processing Transformer) first proposed a pre-trained Transformer for various low-level vision tasks, achieving excellent results on large-scale image restoration benchmarks (Reference 15: Chen, Hanting, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, et al. 2021. “Pre-Trained Image ProcessingTransformer.” *Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, 12299–12310. doi:10.1109 / CVPR46437.2021.01212.). SwinIR utilizes a shift-window self-attention mechanism, surpassing contemporary CNN methods in tasks such as denoising, deblurring, and super-resolution (Reference 16: Liang, Jingyun, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc VanGool, and Radu Timofte. 2021. “SwinIR: Image Restoration Using SwinTransformer.” *Proceedings of the IEEE / CVF International Conference on Computer Vision Workshops (ICCVW)*, 1833–1844. doi:10.1109 / ICCVW54120.2021.00210.).Restormer employs multi-scale transposed attention and gated convolution, significantly reducing the computational complexity of self-attention and achieving optimal performance in weather degradation restoration such as rain and fog removal (Reference 17: Zamir, Syed Waqas, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. 2022. “Restormer: Efficient Transformer for High-Resolution Image Restoration.” *Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, 5728–5739. doi:10.1109 / CVPR52688.2022.00564.). The contribution of this type of method lies in its ability to capture global dependencies through a self-attention mechanism, which is particularly crucial for repairing large-scale missing regions. Its limitation is that the computational complexity is proportional to the square of the image size, resulting in significant resource consumption when processing high-resolution images.

[0065] (4) Methods based on diffusion models

[0066] Diffusion models are a new class of generative models that generate data by progressively adding noise to the data (forward process) and learning to reverse this process (backward process) (Reference 18: Ho, Jonathan, Ajay Jain, and Pieter Abbeel. 2020. “Denoising Diffusion Probabilistic Models.” *Advances in Neural Information Processing Systems (NeurIPS)* 33: 6840–6851.).

[0067] The forward diffusion process gradually diffuses the clear image into Gaussian noise, as shown in Equation (17):

[0068] in, For the initial clear image, For the first The image after adding noise. , Scheduled for a predefined noise variance (a sequence of parameters). It is standard Gaussian noise.

[0069] The reverse denoising process uses a neural network Gradual denoising is used to eventually recover a clear image from random noise. For conditional information (such as degraded images). The optimization objective of this process is usually a simplified variational lower bound, as shown in equation (18):

[0070]

[0071] in, For neural networks that predict noise, the U-Net structure is typically used. In image restoration tasks, the conditional denoising diffusion probability model uses the degraded image as a condition to guide the generation process toward convergence towards a clear image (Reference 19: Saharia, Chitwan, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. 2022. “Image Super-Resolution via IterativeRefinement.” *IEEE Transactions on Pattern Analysis and Machine Intelligence*45 (4): 4713–4726. doi:10.1109 / TPAMI.2022.3204461.). The contribution of this type of method lies in generating high-quality images with good distribution coverage, and producing diverse high-quality restoration results; its limitation is that the inherent iterative sampling process of the diffusion model (usually requiring 50-1000 steps) leads to extremely slow inference speed, making it difficult to meet the needs of real-time applications.

[0072] 3. Single-environment image restoration method

[0073] Researchers have conducted extensive work on the treatment of single degradation types, and have developed a relatively mature technical system.

[0074] Image deraining tasks are mainly divided into two sub-directions: removing rain streaks and removing raindrops. Regarding rain streak removal, Yang et al. designed a multi-scale recurrent network that uses a recurrent structure to progressively extract rain streak features. The core idea is to model the deraining process as... ,in For the rain line residuals predicted by the recursive unit, when t=1 For the input rain image (Reference 20: Yang, Wenhan, Robby T. Tan, Jiashi Feng, Jiaying Liu, ZongmingGuo, and Shuicheng Yan. 2017. “Deep Joint Rain Detection and Removal from a Single Image.” *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)*, 1357–1366. doi:10.1109 / CVPR.2017.183.), Zhang et al. proposed a density-aware network that adaptively adjusts the network depth based on the rain line density. Its density estimation branch outputs a density map. This is used to guide the use of convolutional kernels with different receptive fields in different regions (Reference 21: Zhang, He, Vishwanath Sindagi, and Vishal M. Patel. 2019. “Image De-raining Using a Conditional Generative Adversarial Network.” *IEEE Transactions on Circuits and Systems for Video Technology* 30 (11): 3943–3956. doi:10.1109 / TCSVT.2019.2920407.). Regarding raindrop removal, Qian et al. constructed the first raindrop dataset and designed an attention-based generative adversarial network, whose attention mechanism can be represented as... , This allows the network to focus on the raindrop region (Reference 22: Qian, Rui, Robby T. Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. 2018. “Attentive Generative Adversarial Network for Raindrop Removal from a SingleImage.” *Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, 2482–2491. doi:10.1109 / CVPR.2018.00263.).

[0075] In image desnowing tasks, snowflakes have characteristics such as diverse scales, semi-transparency, and irregular shapes, which make restoration difficult. Liu et al. first constructed the Snow100K dataset and proposed DesnowNet, which achieves end-to-end desnowing through multi-scale feature fusion. The fusion method is shown in Equation (19):

[0076]

[0077] in, For the first Features at each scale For upsampling operation, The fusion weights are learnable (a certain parameter) (Reference 23: Liu, Yun-Fu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. 2018. “DesnowNet: Context-Aware Deep Network for Snow Removal.” *IEEE Transactions on Image Processing* 27 (6): 3064–3073. doi:10.1109 / TIP.2018.2826203.). Chen et al. further analyzed the occlusion effect of snowflakes and proposed a size- and transparency-aware snow removal algorithm, modeling snowflakes as... By estimating transparency To achieve more accurate restoration (Reference 24: Chen, Wei-Ting, Hao-Yu Fang, Jian-Jiun Ding, and Sy-YenKuo. 2021. “PMHLD: Patch Map Based Hybrid Learning DehazeNet for Single ImageHaze Removal.” *IEEE Transactions on Image Processing* 30: 6949–6962. doi:10.1109 / TIP.2021.3097685.).

[0078] 4. Unified Recovery Methods for Complex Environments

[0079] In the face of complex scenarios where multiple weather types coexist in reality, a single degradation model is insufficient to meet application requirements. In recent years, researchers have begun to explore a general recovery framework that can uniformly handle multiple weather degradations.

[0080] The All-in-One model proposed by Li et al. is an early unified multi-weather model. It employs a multi-encoder structure to capture features of different degradation types and fuses them through a shared decoder (Reference 25: Li, Boyun, Yuanbiao Gou, Jiaming Gu, J. Zico Kolter, and M. Salman Asif. 2020. “All-in-One: A UnifiedFramework for Joint Deraining, Dehazing and Desnowing.” *arXiv preprint*, arXiv:2011.12510.). Its output can be represented as... While this method achieved some success, the design of multiple encoders led to a surge in computational costs.

[0081] The TransWeather proposed by Valanarasu et al. employs a more efficient design by introducing learnable weather type queries and intra-regional patch blocks, enabling the network to adaptively handle different types of degradation (Reference 26: Valanarasu, Jeya Maria Jose, Rajeev Yasarla, and Vishal M. Patel. 2022.“TransWeather: Transformer-based Restoration of Images Degraded by AdverseWeather Conditions.” *Proceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision (WACV)*, 1192–1202. doi:10.1109 / WACV51458.2022.00124.). Its core mechanism involves dividing the input image into multiple local patches within the Transformer block and performing self-attention computation within each patch to capture fine-grained features. The entire network operates in a conditionally generated manner, and the output is... ,in This is a learnable weather query vector.

[0082] Zhu et al. proposed a more interpretable method by decomposing features into general weather features and weather-specific features, which are then modeled and fused separately (Reference 27: Zhu, Yurui, Tianrong Zang, Xiaoyu Zhang, and Qiang Wang. 2023. “Generalized Weather Restoration via Disentangled Representation Learning.” *Proceedings of the IEEE / CVFConference on Computer Vision and Pattern Recognition (CVPR)*, 12345–12354.). This decomposition can be formalized as follows: ,in It is a general feature (shared across multiple degradations). Features specific to degradation are selected via a gating mechanism. This method enhances the adaptability to different weather degradations while maintaining the model's lightweight nature.

[0083] In summary, most existing work on the two types of methods mentioned above focuses on simple cases involving a single type of degradation. Research on complex cases involving multiple intertwined degradation factors (such as the simultaneous occurrence of rain, snow, and fog) is relatively limited. This invention primarily addresses this issue, demonstrating good performance in simple cases and showing advantages in complex cases as well. Summary of the Invention

[0084] This invention aims to overcome the difficulties of existing image restoration techniques in complex weather scenarios, such as coarse style feature representation, inability to distinguish fine-grained differences in degradation type and degree, and neglect of information granularity differences in multi-scale feature processing, leading to a difficulty in balancing details and semantics. By proposing an image restoration method based on degradation perception and multi-scale adaptive enhancement, this invention improves the image restoration effect under multiple degradation conditions such as rain, snow, and raindrops, providing a research foundation for computer vision tasks in complex environments.

[0085] This invention primarily focuses on two research perspectives: "degradation-aware style feature learning" and "multi-scale adaptive feature enhancement." Through collaborative optimization, it achieves high-quality reconstruction of degraded images. The overall framework is as follows: Figure 1 As shown.

[0086] The main limitation of existing image restoration theory in dealing with complex weather degradation problems lies in the fact that current methods typically employ a uniform feature processing paradigm, failing to differentiate modeling based on the inherent properties of the features themselves. Specifically, high-resolution regions face different challenges than low-resolution regions; the former suffers from detail loss and edge blurring, while the latter suffers from insufficient global context modeling and limited semantic representation capabilities.

[0087] Based on the above analysis, the overall technical solution proposed in this invention includes: extracting information on the type and degree of degradation to construct fine-grained style features, fusing high and low frequency components of multi-scale features with global-local context, constructing a recovery network model based on conditional modulation, and handling problems such as composite degradation and non-uniform perturbation.

[0088] Step 1: Continuous similarity comparison learning for degradation perception

[0089] The core design objective of this step is to provide downstream recovery networks with fine-grained style features that simultaneously encode information on degradation type and degree, thereby providing more precise conditional guidance.

[0090] This invention proposes a dual-branch style extraction network. Its core idea is to adaptively construct a continuous similarity target based on the differences in degradation type and degree between samples, thereby guiding the ordered distribution of the style feature space. A schematic diagram of the degradation-aware style feature learning module is shown below. Figure 2 As shown.

[0091] The specific implementation includes the following steps:

[0092] (1) Generate degradation probability map and style features

[0093] A dual-scale Transformer encoder is used to extract multi-scale feature maps from the input degraded image. For each scale, its Gram matrix is ​​calculated as shown in Equation (20):

[0094]

[0095] in, For feature maps, Indicates the first and the The correlation between the feature maps of each channel is analyzed. The upper triangular part of the multi-scale Gram matrix is ​​unfolded and concatenated, and after dimensionality reduction by a multilayer perceptron, the style feature vector is obtained. This feature vector serves as conditional information.

[0096] (2) Generate candidate degradation prediction heads

[0097] The concatenated multi-scale features are used as shared features and input into multiple parallel lightweight prediction heads. Each prediction head corresponds to a degradation type (e.g., rain, snow, raindrops), and its structure is a two-layer fully connected network that outputs a scalar. , indicating the degree of prediction for this type of degradation.

[0098] (3) Construction of continuous similarity targets

[0099] For any two samples and Their target similarity The definition is shown in equation (21):

[0100]

[0101] in, and For pre-computed structural similarity labels, and This is the label for the degradation type. This definition ensures that the target similarity of samples of different types is 0, and the target similarity of samples of the same type decreases linearly as the difference in degradation degree increases.

[0102] (4) Region selection and comparison loss calculation

[0103] The weighted mean square error is used to penalize the deviation between the predicted similarity and the target similarity, as shown in Equation (22):

[0104]

[0105] in, For style contrast loss, M is the set of non-self-contrastive sample pairs. and As a stylistic feature, To normalize the cosine similarity, These are the weights for the sample pairs, used to prioritize learning from simpler samples.

[0106] (5) Backpropagation of end-to-end training

[0107] Total loss function As shown in equation (23):

[0108]

[0109] in For the loss of stylistic contrast, To predict the degree of degradation and Mean squared error of the label, For clean samples, the regularization term is... and These are the balancing weights for the corresponding loss terms. In backpropagation, style features... The gradient is directly derived from Dissemination; for guiding similarity targets In Label It does not participate in gradient calculation, but only serves as a fixed supervision signal, thereby stabilizing the contrastive learning process.

[0110] Step 2: Multi-scale adaptive feature enhancement

[0111] This strategy provides differentiated enhancement information to the multi-scale features output by the encoder, guiding the decoder to achieve a balance between detail recovery and structure preservation. The core idea is to adaptively construct a high-resolution scale detail enhancement objective and a low-resolution scale semantic aggregation objective based on the differences in feature resolution and semantic characteristics, guiding an ordered cooperative distribution across feature spaces of different scales. The overall framework of this strategy is as follows: Figure 3 As shown, its internal high- and low-frequency decomposition enhancement branch structure and global-local feature aggregation branch structure are respectively as follows: Figure 4 , Figure 5 Show.

[0112] The specific implementation steps are as follows:

[0113] (1) Division of labor and diversion of multi-scale features

[0114] The four scale features output by the encoder are divided into two groups according to resolution: the high-resolution group (containing the features of the first two layers) , (64 and 128 channels respectively) and low-resolution group (containing the last two layers of features) , (The number of channels is 320 and 512 respectively). High-resolution features are fed into the high- and low-frequency decomposition enhancement branch for detail enhancement, while low-resolution features are fed into the global-local feature aggregation branch for semantic enhancement.

[0115] (2) High- and low-frequency decomposition enhancement at high resolution scale

[0116] For high-resolution input features Perform the following sub-steps:

[0117] High- and low-frequency decomposition: Low-frequency components are extracted through average pooling downsampling. Then, the high-frequency components are obtained by upsampling to the original size using bilinear interpolation and subtracting it from the original features.

[0118] Branch Independent Enhancement: and Two parallel multi-scale dilated convolution branches (with dilation rates of 1, 3, 5, and 7, respectively) are input to obtain the enhanced high-frequency features. and low frequency characteristics .

[0119] Low-frequency attention weighting: for Apply channel attention, as shown in equation (24):

[0120]

[0121] in, This indicates the low-frequency characteristics after channel attention weighting. This indicates element-wise multiplication. This is the Sigmoid function. Then... Upsample to the original size.

[0122] Fusion and Residual Output: Compared with upsampling splicing, through After convolution and SE attention recalibration, compared with the original input The features are then added together to obtain the enhanced high-resolution features.

[0123] (3) Global-local feature aggregation at low resolution scale

[0124] For low-resolution features of the input Perform the following sub-steps:

[0125] Multi-scale local feature extraction: Parallel convolutional layers with dilation rates of 1, 2, 3, and 4 are used to extract multi-scale local features. The concatenated features are then subjected to channel shuffling and... Convolutional fusion yields local features. As shown in equation (25):

[0126]

[0127] in, This indicates a convolutional layer with an expansion rate of k; `cat` represents a concatenation operation along the channel dimension; and `Shuffle` represents a channel shuffling operation. This indicates a 1×1 convolutional layer used to fuse multi-scale features and compress the number of channels.

[0128] Spatial pyramid self-attention: input Flatten as a query The key is obtained through a multi-scale QKV generator (max pooling and flattening concatenation at four scales). Sum Calculate the similarity and perform a weighted sum to obtain the global features. As shown in equation (26):

[0129]

[0130] Will Reshape to the input size and pass through a learnable array initialized to 0. Convolutional processing with residual connections yields global features. .

[0131] Feature fusion: combining local features With global features Element-by-element addition yields the aggregated low-resolution enhanced features.

[0132] (4) Cross-scale collaboration and decoder fusion

[0133] The enhanced features output from steps (2) and (3) are reorganized in their original resolution order and fused across scales through skip connections in the decoder. Specifically, high-resolution enhanced features provide fine edge and texture guidance for the decoder's upsampling process, while low-resolution enhanced features provide global semantic and structural constraints for the decoder. The two types of enhanced features share the same recovery loss gradient during backpropagation, achieving end-to-end collaborative optimization, as shown in Equation (27):

[0134]

[0135] in, To recover the total loss of the branch (the joint loss used for cross-scale collaborative optimization). To rebuild the losses, In order to perceive loss, To balance the weights, this joint loss function allows the gradients of the high- and low-frequency decomposition enhancement branch and the global-local feature aggregation branch to be backpropagated simultaneously, enabling both to adaptively adjust based on the final recovery result and avoiding local optima in a single branch.

[0136] Step 3: Construction and Training of the Joint Recovery Network Model

[0137] (1) Network structure design

[0138] The overall network structure for this step is as follows: Figure 1 As shown, this network aims to extract fine-grained degradation information and multi-scale features from input images to address complex weather degradation problems, and uses this information to guide high-quality reconstruction.

[0139] The network is a dual-branch joint architecture. One branch (style branch) focuses on extracting information about the type and degree of degradation from a global perspective, while the other branch (recovery branch) focuses on differentially enhancing details and semantics from a multi-scale perspective. This module is as follows: Figure 3 As shown.

[0140] The network processing flow, in order of data flow, is as follows:

[0141] First, degrade the image The input is fed into a shared dual-scale Transformer encoder to extract multi-scale features.

[0142] Then, this feature is input into the dual-branch style extraction network described in step one to obtain the style feature vector. and predicted value of degradation On the other hand, this feature is input into a multi-scale adaptive feature enhancement network.

[0143] Finally, the enhanced multi-scale features and style features will be combined. These images are input together into the decoder. The decoder uses upsampling and feature fusion to gradually restore the spatial resolution, ultimately generating the initial restored image. This process uses style features as conditional information to modulate the features of the decoder.

[0144] (2) Network training strategy

[0145] This invention uses paired supervised learning data for training, meaning that each training sample contains a degraded image and a corresponding clean image.

[0146] Data augmentation strategies include: random cropping to Fixed size, random horizontal and vertical flips, and Mixup enhancements.

[0147] The optimization strategy is as follows: use the Adamax optimizer, with an initial learning rate set to... A phased learning rate scheduling strategy is adopted. Specifically, the total number of iterations is 100k, and the learning rate is halved after 50k iterations.

[0148] The final training loss function is obtained by combining the loss functions defined in steps one and two (as shown in equations (23) and (27) respectively), i.e. , as a monitoring signal.

[0149] The training monitoring strategy is as follows: every 500 iterations, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics are calculated on the validation set, and the loss curve and image restoration results are recorded using the TensorBoard tool in order to monitor the training process.

[0150] (3) Summary of overall contributions

[0151] This invention designs a joint network model based on degradation perception and multi-scale adaptive enhancement, belonging to a classic topic in the field of low-level vision and image restoration. Starting from two perspectives—discriminativeness and continuity of feature representation and the differentiated needs of multi-scale processing—this invention comprehensively considers degradation type identification, degradation degree estimation, detail texture preservation, and global structural consistency. It employs continuous similarity contrastive learning and scale-based enhancement strategies, effectively overcoming the poor restoration performance of existing methods in complex weather scenarios, demonstrating significant innovation. Experimental results show that the proposed method outperforms existing mainstream methods in PSNR and SSIM metrics on multiple publicly available synthetic datasets and real-world scenarios, making it more usable in engineering practice. Attached Figure Description

[0152] Figure 1 This is the overall structure diagram of the joint recovery network proposed in this invention, which is described in detail in step three.

[0153] Figure 2 This is a schematic diagram of the structure of the degradation-aware style feature learning module in this invention, which is described in detail in step one.

[0154] Figure 3 This is a schematic diagram of the operation of the multi-scale adaptive feature enhancement module in this invention.

[0155] Figure 4 This is a schematic diagram of the high- and low-frequency decomposition enhancement strategy structure in this invention.

[0156] Figure 5This is a schematic diagram of the global-local feature aggregation strategy structure in this invention. Detailed Implementation

[0157] The image restoration method based on degradation perception and multi-scale adaptive enhancement proposed in this invention will be further described in detail below with reference to the accompanying drawings and experimental embodiments.

[0158] (1) Image dataset and evaluation metrics

[0159] This method uses three publicly available datasets for training and evaluation: Dataset 1 (RainDrop), constructed for the raindrop degradation task, contains images of local raindrop occlusion, with 1129 pairs of images in the training set and 58 pairs in the test set; Dataset 2 (Outdoor-Rain), constructed for the rainline degradation task, contains images of global rainline textures, with 900 pairs in the training set and 300 pairs in the test set; and Dataset 3 (Snow100K), constructed for the snow scene degradation task, contains multi-scale snowflake images, using its Snow100K-L subset, with 10000 pairs in the training set and 1000 pairs in the test set.

[0160] The evaluation metrics used are Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). PSNR is an objective evaluation metric based on pixel mean square error, measured in dB; a higher value indicates better restoration quality. SSIM is an evaluation metric based on image brightness, contrast, and structural information, with a value range of [0,1]; a value closer to 1 indicates greater consistency with the real image.

[0161] (2) Experimental details and main parameter configuration

[0162] Taking the RainDrop dataset as an example, the data consists of pairs of degraded-clean images, divided into training and test sets according to the official partitioning method. Data augmentation includes random cropping... Pixels, random horizontal and vertical flips.

[0163] Model structure parameters: The number of output channels for the dual-scale encoder is 64 and 128, respectively. The dimension of the style feature vector is 64. In the multi-scale adaptive enhancement module, the dilation rate of the dilated convolution of the high-frequency decomposition branch is set to [1,3,5,7], the dilation rate of the global-local aggregation branch is set to [1,2,3,4], and the pooling kernel size of the spatial pyramid pooling is set to [2,3,5,6].

[0164] Optimizer parameters: The Adamax optimizer is used, and the initial learning rate is set to... The momentum parameter uses the default value.

[0165] Learning rate scheduling strategy: A phased decay strategy is adopted, in a total of 100k iterations, the learning rate is halved at 50k steps, and the minimum learning rate is set to... .

[0166] The total number of iterations is 100k, and the batch size is 32. The PSNR / SSIM metric is calculated on the validation set every 500 steps, and the loss value and the recovered image sample are recorded using TensorBoard.

[0167] (3) Experimental results of image restoration task

[0168] The experimental results obtained from the above dataset and parameter configuration are shown in Table 2.

[0169] Table 2. Performance comparison of different methods on three datasets (PSNR / SSIM)

[0170] Art <![CDATA[31.54 / 0.9338 ]]> 29.81 / 0.9088 <![CDATA[30.61 / 0.9083 ]]> <![CDATA[30.65 / 0.9170 ]]> This method 31.89 / 0.9265 30.24 / 0.9111 30.70 / 0.9060 30.94 / 0.9145

[0171] As shown in Table 2, the average PSNR of the method in this invention (30.94) is significantly better than the general method MPRNet and the integrated method TransWeather, and surpasses the recently published Art method. This invention, through joint optimization of degradation-aware style feature learning and multi-scale adaptive enhancement, can simultaneously distinguish degradation type and degree, and adopt different enhancement strategies for high- and low-resolution features. Therefore, it achieves stable and excellent restoration performance in various degradation scenarios such as raindrops, rain lines, and snow scenes. Its cross-scene generalization ability is superior to existing comparative methods, providing solid data support for image restoration tasks in real-world complex environments.

[0172] Table 3. Ablation experimental results of this method on a wider range of datasets.

[0173] Baseline method 29.43 Using only category comparison 30.89 Comparison of usage levels only 30.46 High-frequency-low-frequency enhancement only 30.13 Using only global-local aggregation 30.01 This method (complete) 30.94

[0174] As shown in Table 3, the ablation experiments demonstrate that the proposed degradation-aware contrastive learning module, multi-scale adaptive enhancement module, and joint training strategy are all effective in the final recovery task. Removing any core component leads to performance degradation, validating the indispensability of each module.

[0175] In summary, through comparative and ablation experiments, the proposed method for complex environment image restoration based on degradation perception and multi-scale adaptive enhancement has been fully demonstrated to have significant superiority and effectiveness in removing various degradation types under complex weather conditions.

Claims

1. A method for restoring complex environment images based on degradation perception and multi-scale adaptive enhancement, characterized in that: By extracting style features containing both degradation type and degree information from the input degraded image, precise conditional guidance is provided for the restoration process. Then, the multi-scale features output by the encoder are differentially enhanced according to resolution. High-resolution features are enhanced with detailed texture through high- and low-frequency decomposition, while low-resolution features are enhanced with semantic representation through global-local aggregation. Finally, the enhanced multi-scale features and fine-grained style features are input into the decoder, and a high-quality restored image is generated through conditional modulation. The implementation steps are as follows: S1, Continuous similarity comparison learning for degradation perception This step is used to extract fine-grained style features from the input degraded image that simultaneously contain information on the type and degree of degradation, providing precise conditional guidance for the downstream restoration network; specifically, it includes the following operations: First, a dual-scale Transformer encoder is used to extract multi-scale feature maps from the input degraded image. For each scale, its Gram matrix is ​​calculated as shown in Equation (1): in, Representing feature maps, Indicates the number of channels. Indicates altitude, Indicates width, Indicates the first The first channel and the first Correlation between feature maps of each channel; Then, the upper triangular portion of the multi-scale Gram matrix is ​​unfolded and concatenated, and the style feature vector is obtained after dimensionality reduction using a multilayer perceptron. Simultaneously, the concatenated multi-scale features are input as shared features into multiple parallel lightweight prediction heads, each corresponding to a degradation type, and outputting a scalar. , indicating the degree of prediction for this type of degradation; Subsequently, for any two samples i and j, a continuous similarity target is constructed as shown in equation (2): in, Indicates target similarity. and This indicates pre-computed structural similarity labels. and The label represents the degradation type. This definition ensures that the target similarity of samples of different types is 0, and the target similarity of samples of the same type decreases linearly as the difference in degradation degree increases. Finally, the style contrast loss is calculated using the weighted mean squared error method. As shown in equation (3): in, This represents the set of non-self-comparison sample pairs. and Indicates stylistic characteristics, Represents the normalized cosine similarity function. This represents the weights of sample pairs, used to prioritize learning from simpler samples. This represents the target similarity; this step is optimized through end-to-end backpropagation, and the total loss function is... As shown in equation (4): in, For the loss of stylistic contrast, This represents the mean squared error of the predicted degradation level and structural similarity label. The regularization term representing clean samples. and This step represents the balance weights; it yields fine-grained style features, providing conditional information for the subsequent restoration process to distinguish the type and degree of degradation. S2, Multi-scale Adaptive Feature Enhancement This step is used to provide differentiated enhancement information to the multi-scale features output by the encoder according to their resolution, so as to guide the decoder to achieve a balance between detail recovery and structure preservation; specifically, it includes the following operations: (1) Division of labor and diversion of multi-scale features The four scale features output by the encoder are divided into two groups according to resolution: the high-resolution group contains the first two layers of features, and the low-resolution group contains the last two layers of features; the high-resolution features are sent to the high-low frequency decomposition enhancement branch for detail enhancement, and the low-resolution features are sent to the global-local feature aggregation branch for semantic enhancement. (2) High- and low-frequency decomposition enhancement at high resolution scale For high-resolution features High- and low-frequency decomposition enhancement is performed: low-frequency components are extracted through average pooling downsampling. Then, the high-frequency components are obtained by upsampling to the original size using bilinear interpolation and subtracting it from the original features. ;Will and The enhanced high-frequency features are obtained by inputting two parallel multi-scale dilated convolution branches respectively. and low frequency characteristics ;right Channel attention is applied as shown in equation (5): in, This indicates the low-frequency characteristics after channel attention weighting. This indicates element-wise multiplication. This represents the Sigmoid function. This represents a multilayer perceptron. Indicates average pooling. This indicates max pooling; subsequently, Upsampled to the original size, and After splicing Convolution and attention recalibration, then compared with the original input. Adding them together yields the enhanced high-resolution features; (3) Global-local feature aggregation at low resolution scale For low-resolution features Perform global-local feature aggregation: use parallel convolutional layers with different dilation rates to extract multi-scale local features, concatenate them, and then perform channel shuffling. Convolutional fusion yields local features. As shown in equation (6): in, This represents a convolutional layer with an inflation rate of k. express convolution, This indicates a channel mixing operation. This indicates a splicing operation. Indicates the expansion rate The convolution is performed; at the same time, the input is flattened as a query, and the key and value are obtained through a multi-scale key-value generator. The spatial pyramid self-attention is then calculated to obtain the global features as shown in Equation (7): Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d represents the channel dimension; Global features are obtained by reshaping G to the input size and passing it through learnable residual connections. Then local features With global features Element-wise addition yields the aggregated low-resolution enhanced features; Finally, the enhanced features output from the high- and low-frequency decomposition enhancement branch and the global-local feature aggregation branch are reorganized in the original resolution order, and cross-scale fusion is performed through the skip connections of the decoder. The joint loss function shown in Equation (8) is used to achieve end-to-end collaborative optimization. in, To recover the total loss of the branch (the joint loss used for cross-scale collaborative optimization). Indicates the losses incurred during reconstruction. Indicates perceived loss. This represents the balanced weights; through this step, high-resolution enhancement features provide the decoder with fine edge and texture guidance, while low-resolution enhancement features provide the decoder with global semantic and structural constraints. S3. Construction and Training of Joint Recovery Network Model This step involves constructing a dual-branch joint architecture network model and achieving high-quality reconstruction of degraded images through end-to-end training; specifically, it includes the following operations: First, a joint restoration network model is constructed: the degraded image is input into a shared dual-scale Transformer encoder to extract multi-scale features; these features are then input into the dual-branch style extraction network described in step S1 to obtain style feature vectors. and predicted value of degradation On the other hand, the input is fed into the multi-scale adaptive feature enhancement network described in step S2; then the enhanced multi-scale features are combined with style features. The images are input into the decoder, which gradually restores the spatial resolution through upsampling and feature fusion. Style features are used as conditional information to modulate the decoder features, and finally, the restored image is generated. Then, training is performed using paired supervised learning data, with each training sample containing a degraded image and its corresponding clean image; data augmentation strategies include random cropping to a fixed size, random horizontal and vertical flipping, and hybrid augmentation; the optimization strategy employs an adaptive moment estimation optimizer and a phased learning rate scheduling strategy; the final training loss function is... ,in As shown in equation (4) in step S1, As shown in equation (8) in step S2; Finally, a training monitoring strategy is adopted to calculate and record evaluation metrics on the validation set during the training process in order to monitor the training process. Through this step, the joint optimization of degradation perception and multi-scale adaptive enhancement is achieved, and high-quality image restoration is completed under complex weather conditions.

Images