Bidirectional decoupling translation unsupervised defogging network with feature and pixel comparative representation

Abstract

The invention provides a bidirectional decoupling translation unsupervised defogging network with feature and pixel comparison representation, decoupling modeling of haze and clear content in an image is realized through a feature and pixel comparison mechanism, and paired clear images do not need to be supervised and trained. The whole network is composed of five key modules: 1) a content recovery branch adopts a multi-scale residual structure to recover image details and textures, and cooperates with a semantic enhancement and color correction module to improve visual quality; 2) the parameter estimation branch estimates a transmissivity map and atmospheric light based on an atmospheric scattering model, and improves the prediction accuracy through context modeling; 3) a feature comparison learning module strengthens the discrimination of clear and fog features; 4) introducing closed-loop consistency by the reverse fogging structure based on reverse physical modeling, and providing an unsupervised training signal; and 5) a pixel comparison mechanism is used for constructing a comparison path by using the direct / indirect defogged image, the reverse fog-added image and the original image, and the network discrimination capability and generalization performance are enhanced. The experiment is carried out on a real natural fog image data set, the result shows that the proposed method is superior to the existing defogging method in the aspects of structural fidelity, detail reduction and color consistency, and the practicability of the method in a complex environment is verified.

Description

Technical Field

[0001] It mainly involves technical fields such as image dehazing, unsupervised learning and feature decoupling. By utilizing non-matching image data, a bidirectional decoupled translation unsupervised dehazing network with feature and pixel contrast representation is proposed. Background Art

[0002] With the rapid development of modern industrial activities, environmental pollution has become increasingly prominent, and the negative impact of haze on people's daily lives has become increasingly severe. In particular, during image acquisition, heavy haze can significantly degrade image quality, causing blur, color distortion, and loss of detail. This not only affects human visual perception but also severely impairs the performance of advanced visual tasks such as object detection and image recognition. To address this, image dehazing technology has emerged, aiming to restore clear, realistic image content from hazy images. It has become a key research topic attracting widespread attention in academia and industry.

[0003] Previous research has typically used an atmospheric scattering model to describe the effects of haze on image formation. This model primarily includes two key parameters: the atmospheric scattering coefficient and transmittance, which control the extent of haze-induced image degradation. Based on this model, early image dehazing methods relied on prior knowledge of the image. By designing appropriate priors, they estimated these two intermediate parameters, thereby restoring a clear, haze-free image. However, the image priors relied on by these methods often deviate from the real scene, leading to inaccurate estimation of the intermediate parameters and, in turn, compromising the final dehazing performance. With the rapid development of convolutional neural networks, a growing number of studies have shifted towards data-driven image dehazing methods. These methods typically leverage the powerful feature representation and fitting capabilities of convolutional neural networks to predict intermediate parameters or directly restore the clear image in an end-to-end manner, achieving significant performance improvements over traditional prior-based methods. However, these learning-based methods typically rely on a large number of synthetic hazy / clear image pairs for training. However, obtaining real paired images is expensive and difficult in practice. This leads to domain differences between synthetic and real data during training, which in turn affects the model's generalization performance in real-world scenarios. Therefore, how to effectively use unpaired training images for network optimization and improve the dehazing effect of the model in real scenes has become a key challenge in current image dehazing research and an important direction for promoting further development in this field. Summary of the Invention

[0004] To address the challenges of existing image dehazing methods in unsupervised environments, such as inaccurate content restoration, insufficient separation of haze information, and insufficient ability to recover image details, a novel bidirectional disentangled translation network architecture is proposed for high-quality unsupervised image dehazing. Unlike traditional approaches that treat image dehazing as an image enhancement or direct image mapping problem, this approach is modeled as a feature disentanglement task, explicitly separating haze-related factors from the image's intrinsic content while preserving the original image structure. To this end, a two-branch network architecture is designed, consisting of a content restoration branch that extracts structurally stable and semantically clear image content from the haze image, emphasizing the preservation of image details and edges; and a parameter estimation branch that specifically extracts haze-related low-frequency features in the image and predicts key parameters in the atmospheric scattering model (including transmittance maps and global atmospheric light). To achieve effective feature separation between these two branches, a multi-level feature contrast representation mechanism is introduced. This mechanism comprehensively considers the self-similarity of features within the same branch and the mutual exclusivity of features between the two branches, significantly improving the separability of content and haze information in the feature space. In addition, to further enhance the model's constraint capability under unsupervised conditions, a closed-loop system integrating forward dehazing and reverse rehazing processes was constructed. Specifically, a classic atmospheric scattering physics model is embedded in the system, enabling the network to: utilize the atmospheric light and transmittance maps output from the parameter estimation branch to achieve forward image restoration from a foggy image to a clear image; then, using the dehazed image output by the content recovery branch, the corresponding rehazed image is generated through the above-mentioned physical model to form a physically reversible image transformation closed loop, thereby constructing an unsupervised cycle consistency constraint and training the network without the need for paired clear images. At the same time, a bidirectional pixel-level contrast learning mechanism was designed, which constructs multiple positive and negative contrast pairs between the original foggy image, the content recovery map, the parameter estimation map, and the heavy foggy image. Specifically, the system guides the network to more accurately distinguish image representations with different semantics and haze characteristics by bringing pixels of similar images closer together and pixels of heterogeneous images farther apart, effectively alleviating the training perturbations and uncertainties caused by different data domain distributions.

[0005] In summary, the proposed bidirectional decoupled image dehazing system is significantly innovative and practical in terms of architecture design, feature modeling, physical prior fusion, and contrast constraint mechanism. It can achieve robust image dehazing effects with rich detail preservation without the need for paired training data, and is widely applicable to multiple practical scenarios such as security monitoring, autonomous driving, and remote sensing image processing.

[0006] The technical solution adopted is a bidirectional decoupled translation unsupervised dehazing network with feature and pixel contrast representation. It is implemented in the following steps:

[0007] Step 1: Design of an overall unsupervised image dehazing framework. We propose an unsupervised image dehazing framework based on feature decoupling. This method aims to address the problems of existing supervised dehazing methods, such as their reliance on real blurry / clear image pairs and weak generalization capabilities. We innovatively reconstruct the image dehazing task into a feature-level decoupling modeling problem. The dehazing framework mainly includes the following core modules:

[0008] First, a two-branch collaborative learning mechanism is designed to effectively separate and reconstruct different semantic information in the input fog image. Specifically, the content recovery branch is used to extract structured and clear semantic content information in the image, focusing on restoring the original details and texture structure of the scene; the parameter estimation branch is used to extract imaging degradation characteristic parameters that characterize the haze components and model the impact of haze on image formation.

[0009] Secondly, to further improve the decoupling performance of the two feature branches mentioned above, a feature contrast representation mechanism is introduced. This mechanism strengthens the feature differences between the image semantic content and haze information in the latent representation space by constructing positive and negative sample pairs, thereby enhancing the discriminative ability and separability of the feature representation. In addition, considering the dual relationship between image dehazing and image hazing in the physical imaging process, a cycle consistency mechanism based on forward mapping and backward mapping is constructed. This mechanism combines a physical degradation model as a guide to establish a closed-loop reconstruction path between the input image and the restored image, thereby achieving effective model optimization and constraint in unsupervised learning scenarios.

[0010] To further improve image restoration quality, a pixel-by-pixel contrast constraint strategy is proposed, introducing semantic and structural alignment supervision signals at the image representation level to enhance the network's ability to model local details and global consistency, ensuring that dehazed images retain clarity while retaining good content restoration and naturalness. Through a series of innovative designs, including feature decoupling, contrast representation, cycle consistency, and pixel-level structural constraints, an end-to-end unsupervised image dehazing learning framework has been constructed. This framework, independent of paired image data, achieves high-quality image dehazing in complex, realistic environments, promising broad application value and widespread adoption.

[0011] Step 2: The content restoration branch is designed to efficiently extract the underlying clear image structure information from the input foggy image, providing structured semantic support for the final image dehazing result. To achieve this goal, a set of clear content feature mining modules is proposed and constructed, mainly including the following key submodules:

[0012] Multi-level convolutional encoder: A multi-level feature encoding module for layer-by-layer encoding of input images is proposed. This module effectively extracts the structural information of the image at different levels through layer-by-layer convolution stacking and feature downsampling mechanism, including low-level texture features, mid-level edge contours, and high-level semantic representations, to achieve comprehensive modeling of the multi-scale spatial structure of the image. Specifically, this module adopts a hierarchical progressive encoding structure, combined with a multi-scale receptive field design strategy, to perform deep feature extraction on the input image. The shallow network focuses on capturing details and texture information, and is suitable for restoring tiny structures in the image that are obscured by fog; the middle layer strengthens the structural consistency of areas such as edges and contours; the deep network is responsible for extracting global semantic features to assist in judging the content and layout of the main area of ​​the image. The above-mentioned hierarchical features are integrated across scales in the subsequent fusion module to further enhance the network's ability to express image spatial features and the accuracy of discrimination;

[0013] Residual Feature Extractor: To effectively alleviate the vanishing gradient problem that occurs during the training of deep neural networks and improve the continuous expression of image structural information and detailed features, a residual connection mechanism is introduced into the feature extraction and recovery module. This mechanism constructs an identity mapping path between different layers, directly transferring shallow features to deep layers, effectively maintaining the integrity and stability of feature flow, and significantly improving the training efficiency and convergence ability of the model;

[0014] Semantic enhancement unit: To enhance the semantic information expression capability of key areas in haze images (such as the main image contour, edge structure, etc.), the present invention introduces an attention mechanism and a context aggregation structure in the feature extraction and fusion process to construct a key area perception enhancement module. This module automatically focuses on semantically important areas in the image through the joint modeling of spatial attention and channel attention, effectively improving the model's ability to identify and restore key area features. In addition, combined with the context information aggregation strategy, the global and local structural relationships are captured through a multi-scale context-aware path, enhancing the expression capability of edge and detail features while retaining global semantic consistency;

[0015] Frequency enhancement module design: In view of the fact that low-frequency components dominate in haze images and image details are easily obscured, a frequency domain guidance mechanism is introduced in the feature extraction stage to construct a frequency enhancement module. This module separates low-frequency and high-frequency components by performing frequency domain transformation on the input features, and uses low-frequency information suppression filtering and high-frequency detail enhancement operators to effectively enhance high-frequency features such as edges and textures of the image. Specifically, the module first extracts the spectral characteristics of the image in the frequency domain, applies a learnable filter to suppress flat and redundant low-frequency areas, and at the same time enhances the response amplitude of the high-frequency channel, thereby improving the model's sensitivity and discrimination ability to detail information;

[0016] Adaptive Color Correction Module: This module introduces a multi-scale convolution kernel structure to fully extract and restore image details and structural information at different scales, effectively improving image clarity and layering. Furthermore, the module dynamically adjusts hue and color mapping parameters based on the color distribution characteristics of the input image to accurately compensate for color shifts caused by haze. This module possesses self-learning capabilities and can adaptively adjust correction strategies based on image content, achieving more natural and realistic color reproduction, significantly enhancing the visual realism and perceptual quality of dehazed images.

[0017] In the forward propagation stage of the network, the content recovery branch takes the original fogged image as input, relies on the above modules to jointly extract effective features, and directly generates a preliminary restored clear image result. This image not only serves as one of the final dehazing outputs, but also serves as the basic input for subsequent image contrast learning and inverse haze synthesis modules, establishing an important linkage interface in the "dehazing-adding fog" closed-loop process.

[0018] Step 3: Designing the parameter estimation branch. This branch proposes a feature extraction path for modeling physical image degradation. Its core task is to accurately estimate key physical parameters of the degradation process, including atmospheric light values ​​and transmittance maps, for the input haze image based on a classical atmospheric scattering model. To enhance the accuracy and robustness of parameter estimation, this branch combines two structural modules: a global context-aware module and a local detail estimation module.

[0019] Design of global context perception module: In order to improve the model's ability to understand the overall illumination structure of the image, the present invention designs a global context perception module, which performs context modeling on the input image based on the multi-scale pooling mechanism and the global attention mechanism, and comprehensively extracts the global brightness distribution characteristics and scene illumination change rules in the image. Specifically, the module captures the brightness contrast and structural relationship of the image in the global and local range by applying pooling operations of different scales in the spatial dimension; at the same time, combined with the self-attention mechanism, it models the illumination dependency between different regions, thereby achieving accurate positioning of potential atmospheric light areas and estimation of illumination values. This module effectively alleviates the atmospheric light estimation error caused by uneven fog concentration or occlusion in local areas, and improves the stability and consistency of atmospheric light value prediction in the overall defogging process. By enhancing the understanding of global contextual information, the module further improves the coordination and naturalness of brightness correction, color restoration and structure reconstruction in the image defogging process;

[0020] Local Detail Estimation Module Design: This paper proposes a local detail estimation module for enhancing the accuracy of transmittance modeling in local image regions. This module, built on a multi-layer convolutional structure, is highly sensitive to detailed areas such as edges and textures. By locally modeling and enhancing the response of the image feature map in the spatial dimension, the accuracy and robustness of transmittance estimation in key areas are significantly improved. Specifically, the module introduces multiple convolution kernels with different receptive fields during the feature extraction stage to enhance the ability to capture local image degradation features (such as edge blur and texture attenuation). By guiding the network to focus on areas of severe image degradation or structural mutations, it effectively compensates for the distortion and loss of local information caused by haze. At the output stage, the module generates high-resolution, high-precision local transmittance estimates, providing accurate prior support for subsequent atmospheric light correction and image restoration. The introduction of this module not only improves the accuracy of the transmittance map in the spatial dimension but also enhances the overall dehazing system's ability to recover detail.

[0021] To further leverage the predicted physical parameters, the input image is reconstructed using a physical model guided by atmospheric scattering imaging formulas, generating an indirect dehazed image based on physical reasoning. This image serves not only as an auxiliary output from the physical path but also as a structural and color comparison supervision signal for the direct output image of the content recovery branch. This creates a dual-path fusion mechanism of data-driven and physical modeling for the dehazing task, effectively improving the realism and structural integrity of the final image.

[0022] Step 4: Design a feature contrast representation mechanism. To enhance the decoupling capabilities of different information flows in the latent feature space, a joint contrast representation mechanism is proposed. During the training phase, this mechanism constructs positive and negative sample pairs to encourage similar features to aggregate in a high-dimensional space and separate dissimilar features. Specifically, this mechanism maintains the self-similarity of features within the same information flow and strengthens the mutual exclusivity of features between different branches to improve the separation of content information from haze components. This design not only enhances the model's ability to discriminate image components but also provides a supportive representation space for subsequent image consistency constraints.

[0023] Step 5: Design of the reverse fogging architecture. To achieve complete closed-loop learning in unsupervised scenarios, a reverse fogging module is introduced. This module takes the clean image generated by the content recovery branch as input and combines it with the atmospheric light and transmittance information predicted by the parameter estimation branch to reconstruct a reverse fogged image. This image is compared with the original input fog image, and a cycle consistency loss constraint is established. This allows the model to achieve end-to-end optimization during unsupervised training, ensuring that image content and structure are preserved during the conversion process.

[0024] Step 6: Design of a pixel-level contrastive representation strategy. To further enhance the consistency and differentiation of the results across image paths, a pixel-level contrastive representation strategy was designed. This strategy constructs a set of multi-dimensional pixel representation contrast constraints around four image types: original hazy images, directly dehazed images, indirectly dehazed images, and reversely hazed images. This strategy performs the following operations in pixel space: it brings pixel features from the same semantic category or visual structure closer together, while also separating pixel features from different image states in the representation space. This improves the quality of the final image restoration and significantly reduces interference between data domains.

[0025] The present invention is also characterized in that:

[0026] Step 1 Implementation: Within the overall framework, image dehazing is modeled as a feature decoupling process, separating the mixed features in the input fog image into two distinct information factors: a clean factor representing the true image structure, and a fuzzy factor representing the interference characteristics of the haze. In practice, a two-branch decoupling structure is constructed to extract content features related to image semantics and haze features related to degradation, respectively, and these two types of features are represented separately in the latent space. The system then differentially models and processes these two types of features, effectively reconstructing a clear image while avoiding the risks of overfitting or information aliasing common in traditional methods.

[0027] Step 2 Implementation: In the design of the content recovery branch, potential clear image structural information is efficiently extracted from the input fog image, providing structured semantic support for image dehazing. This branch integrates multiple modules working together, including a multi-level convolutional encoder to extract multi-scale texture and semantic information, a residual feature extractor to enhance edge and detail representation and prevent gradient vanishing, a semantic enhancement unit to highlight key regional features using an attention mechanism, a frequency enhancement module to improve dehazing robustness by separating high- and low-frequency information, and an adaptive color correction module to dynamically adjust hue to improve color cast. The final output of the preliminary clear image serves as both the dehazing result and an important input for contrastive learning and haze synthesis, forming a closed-loop "dehazing-adding fog" structure.

[0028] Step 3 Implementation: In the parameter estimation branch, to accurately model the haze degradation process in the image, an inverted triangle network structure is adopted to extract haze-related features from top to bottom through a wide input and deep convergence approach. Considering the uneven distribution and varying concentrations of haze in different spatial locations and image regions, this branch utilizes a multi-scale modeling mechanism, treating haze as a superposition effect of a mixture of multiple scales and concentrations. To this end, the following two modules are introduced: 1) Global Context Perception Module: This module effectively models the global brightness distribution and scene illumination changes in the image through multi-scale pooling or global attention mechanisms, improving the stability and consistency of atmospheric light value prediction; 2) Local Detail Estimation Module: This submodule uses a convolutional structure to enhance the transmittance modeling capability of local regions, paying particular attention to the degradation feature changes in edge and texture regions, thereby improving the spatial accuracy of the transmittance map.

[0029] Step 4 Implementation: In the feature contrast representation module, a cross-domain contrastive learning mechanism based on decoupled features is designed. Specifically, the encoded feature representation of the input image is decomposed into two subspaces: one representing structurally stable, clear content features, and the other representing fuzzy, hazy features due to environmental interference. To enhance feature discrimination and separation, a contrastive representation loss function is established by leveraging the similarity (e.g., semantic consistency) and mutual exclusivity (e.g., differences in clarity and blur) between features. This strategy not only enhances the network's ability to distinguish between different feature components (foggy and clear information) but also improves the efficiency of filtering redundant information in the input image.

[0030] Step 5: Implementation of the reverse fogging module: The direct dehazed image output by the content restoration branch is used as input. Combined with the atmospheric scattering coefficient and transmittance maps predicted by the parameter estimation branch, the haze image is reconstructed to form a reverse fogging path. This process strictly simulates the image according to the physical model of atmospheric scattering to maximize the authenticity of the fogging process. The generated reverse fogged image is constrained with the original input fog image using a cycle consistency loss, ensuring that the model maintains semantic structure and texture consistency between dehazed and heavy fog, improving the network's robustness and generalization capabilities.

[0031] Step 6 Implementation: In the pixel-level contrast representation module, four key image results are selected: original foggy images, direct dehazed images, indirect dehazed images based on physical parameters, and reverse fogged images. A contrastive learning mechanism is then constructed in the pixel representation space. Direct and indirect dehazed images are considered as positive sample pairs, forcing them to be close in space, while they form negative sample pairs with the original foggy image, increasing their distance in the representation space. On the other hand, reverse fogged images are defined as sample types that should be closer to the original foggy image and further away from the clear image. This contrast strategy significantly improves the quality of image representation while strengthening the mutual constraints between multi-path outputs, helping to generate more stable and natural image dehazing results.

[0032] In summary, the main technical contributions of the present invention are:

[0033] (1) A bidirectional decoupled translation-based unsupervised image dehazing network based on feature and pixel contrast representation is constructed. A dual-path feature decoupling mechanism is introduced to branch model the image content and haze information. A feature-level and pixel-level contrast representation strategy is jointly designed. Combined with a cycle consistency constraint mechanism, high-quality image dehazing can be achieved without relying on paired haze-clear image training data.

[0034] (2) An unsupervised network structure is proposed to model image dehazing as a feature disentanglement task. In the proposed decoupling framework, the clear content information and the blurred haze information in the image are processed by the content recovery branch and the parameter estimation branch respectively, thereby improving the target clarity and semantic accuracy of haze removal and effectively avoiding information confusion and feature loss.

[0035] (3) An efficient feature contrast representation mechanism is designed to explicitly separate haze features from image content in the latent space. This mechanism considers the self-similarity within each information stream and introduces the mutual exclusivity modeling between different features, thereby improving the distribution discrimination ability in the feature space and helping to improve the generalization and robustness of the dehazing network.

[0036] (4) The forward dehazing and reverse hazing processes are integrated into the same network framework, and pixel contrast representation is introduced for fine-grained constraints. By constructing a closed-loop structure from content recovery to reverse hazing, multiple constraints on image clarity, content consistency, and physical reversibility are implemented. At the same time, pixel contrast representation is used to alleviate the learning confusion problem caused by different data distributions, significantly enhancing the network's dehazing performance in unsupervised scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the overall structure of the proposed bidirectional disentangled translation unsupervised dehazing network with feature and pixel contrast representation;

[0038] Figure 2 Schematic diagram of the training, evaluation and testing process of the proposed bidirectional disentangled translation unsupervised dehazing network with feature and pixel contrast representation;

[0039] Figure 3 Visual comparison results of the proposed bidirectional disentangled translation unsupervised dehazing network with feature and pixel contrast representation and other dehazing methods. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the described embodiments are intended only to facilitate understanding of the present invention and do not limit it in any way. The accompanying drawings are in a very simplified form and are not in precise proportions. They are only used to conveniently and clearly assist in illustrating the embodiments of the present invention. The structures shown in the accompanying drawings are part of the actual structure. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] We propose a bidirectional decoupled translation unsupervised dehazing network with feature and pixel contrast representations. We consider image dehazing as a feature decoupling task. By separating the clear content-related information from the haze-related feature elements in the original hazy input image, we can achieve image dehazing without the need for a matched haze-clear pixel dataset.

[0042] As attached Figure 1 As shown in Figure 1, the overall structure of an unsupervised image dehazing network is provided, which mainly includes the following five key components:

[0043] (1) Content restoration branch design: This branch is mainly used to directly restore a clear image from the input fog image. By constructing a multi-scale feature extraction module and a residual connection structure, the details and texture information in the image are gradually reconstructed, thereby achieving end-to-end unsupervised image restoration.

[0044] (2) Parameter estimation branch design: This branch is based on the classic atmospheric scattering physical model and predicts key parameters in the image degradation process, including the transmittance map and the global atmospheric light. After completing the parameter estimation, the image can be modeled and restored using the physical defogging formula to obtain physically consistent defogging results;

[0045] (3) Feature contrast representation design: This module effectively improves the discriminability and robustness of the extracted features by introducing a feature-level contrast learning mechanism. The basic idea is to shorten the representation distance between different views of the same image in the feature space and distance the representation distance between different images, thereby improving the network's adaptability to diverse scenes and complex haze conditions;

[0046] (4) Inverse fogging structure design: To enhance the structural constraint capability of the network, an inverse fogging mechanism is further introduced. Based on the atmospheric scattering model, this mechanism re-fogs the restored clear image to generate a pseudo-fog image, and establishes consistency constraints with the original input image in the pixel space, thereby guiding the network to learn the mapping relationship between the fog image and the defogged image;

[0047] (5) Pixel contrast representation design: A contrast learning strategy is introduced at the pixel level. By constructing the representation differences between different pixel pairs in the image, the network's recovery ability in edge, detail and texture areas is enhanced, thereby improving the fineness and consistency of image restoration.

[0048] Ultimately, through the collaborative work of the above five modules, the constructed network framework can output two independent image dehazing results: a dehazed image directly generated by the content recovery branch, which has stronger detail expression capabilities; and an indirect dehazed image calculated by the parameter estimation branch combined with the physical model, which has higher physical consistency and interpretability.

[0049] Furthermore, the inverse haze module constructed in the framework is based on the direct dehazed image and combines the estimated transmittance map and atmospheric light parameters to regenerate the haze image; a cycle consistency constraint in pixel space is established between the pseudo haze image and the original input haze image, thereby providing a pseudo-contrast supervised training signal under unsupervised learning conditions, significantly enhancing the stability and convergence of unsupervised model training, and achieving better dehazing performance.

[0050] In the content restoration branch, the core goal is to explicitly restore the clear content information in the image, so the construction of the feature extraction module is particularly critical. As the foundational layer of the entire network structure, this branch undertakes the core task of removing interference from complex foggy images and restoring the original clear scene.

[0051] The multi-level convolutional encoder constructs a multi-layer convolutional network from shallow to deep, encoding the features of the input image step by step, extracting information of different semantic granularity at different levels. Shallow convolution focuses on the basic texture and color distribution in the image, capturing low-level details such as lines and textures; mid-level convolution focuses on moderately complex features such as edges and structures, helping to depict the outlines and structural relationships of objects in the image; deep convolution extracts global contextual semantic information, improving the model's understanding and representation of the entire scene.

[0052] To address the common problems of vanishing gradients and training degradation in deep networks, the residual feature extractor uses a classic residual learning mechanism, introducing short-circuit connections between convolutional layers to achieve information flow and difference enhancement between input and output features. This structure not only maintains the stable conduction of gradients during backpropagation, but also effectively improves the network's nonlinear expression capabilities. In image dehazing tasks, this module has a significant repair effect on problems such as blurred edges and texture fractures in foggy areas. In particular, it can enhance the ability to capture structural continuity in images, making the restored image sharper and more clearly defined, while retaining more detailed textures, providing high-quality structural information support for image restoration.

[0053] The semantic enhancement unit aims to improve the network's ability to focus on key semantic areas. It uses a combination of spatial attention mechanisms, channel attention mechanisms, and contextual feature aggregation strategies to achieve deep mining and precise focusing of image semantic features. By assigning differentiated weights to the importance of different spatial locations or channel dimensions, the network can actively "focus" on the areas in the fog image that are most critical for perception and structural recovery, such as subject boundaries, foreground outlines, and target center areas. By integrating multi-scale contextual information, the ability to model the semantic relationships of occluded or blurred areas is further improved, thereby enhancing the recognition and restoration performance of key areas in the restored image, and achieving the transformation from rough images to highly semantically clear images.

[0054] Given that haze primarily attenuates and blurs the low-frequency energy of images, the design incorporates a low-frequency feature suppression module and a high-frequency feature enhancement mechanism to improve the model's selective focus on frequency domain information and enhance its ability to recover image detail and texture. The low-frequency feature suppression module effectively suppresses low-frequency interference caused by haze, reducing the impact of blurring information; while the high-frequency enhancement mechanism enhances high-frequency details such as edges and textures, improving image sharpness and clarity.

[0055] To further enhance the diversity and richness of content expression, the content recovery branch employs a multi-scale convolutional kernel structure, fusing receptive fields of different scales to capture information at different levels of the image. Small-scale convolutional kernels help extract detailed features (such as texture and edges), while large-scale convolutional kernels focus on modeling overall structure and contextual information, enabling more comprehensive content recovery.

[0056] To effectively address the color shift caused by haze interference, an adaptive color correction module was designed. This module learns the color distribution characteristics of the input haze image, dynamically generates a color mapping function, and adaptively adjusts the image hue. This makes the color of the dehazed result closer to that of a realistic, clear scene, improving visual quality and naturalness.

[0057] In the parameter estimation branch, the network's core goal is to accurately model and estimate haze-related physical parameters in the input image, primarily the atmospheric scattering coefficient and transmittance map. Starting from a physical model of image degradation, this branch uses deep feature learning to extract low- and mid-level information reflecting haze distribution intensity and component concentrations. Furthermore, combined with contextual environmental perception, it performs spatially adaptive inference on transmittance and atmospheric light distribution across different regions.

[0058] This network employs an inverted triangle feature extraction architecture in its parameter estimation branch. It consists of a shared encoder and two independent decoders, enabling detailed modeling and prediction of physical properties (atmospheric light parameters and transmittance maps). The encoder first performs feature upscaling and compression on the input haze image, extracting cross-scale spatial and semantic features to provide high-quality feature support for subsequent parameter estimation. The two decoders, respectively, perform decoupled modeling of atmospheric light and transmittance, using parallel regression to improve the accuracy and stability of parameter estimation.

[0059] To further enhance the parameter estimation branch's adaptability to complex haze environments and the accuracy of physical parameter prediction, the network design introduces two key auxiliary modules to enhance the modeling capabilities of atmospheric light values ​​and transmittance maps, respectively: a global context perception module and a local detail estimation module.

[0060] The global context perception module aims to model global illumination and brightness distribution information in an image. By introducing multi-scale pooling structures (such as pyramid pooling and dilated convolution) or self-attention mechanisms, it achieves perception and integration of the image's overall lighting environment. In particular, when scenes contain strong light source variations or deep depth of field, this module effectively captures the influence of atmospheric light in different spatial regions, thereby avoiding local deviations or regional inconsistencies in atmospheric light estimation. By modeling the aggregation of global features, the stability, continuity, and physical consistency of atmospheric light value predictions are significantly improved, providing reliable guarantees for subsequent brightness restoration of dehazed images.

[0061] To accurately characterize the spatial distribution of transmittance in an image, the local detail estimation module focuses on modeling detailed information in local regions. By using a deep convolutional structure to enhance the modeling of high-frequency regions such as image edges, textures, and contours, it can more sensitively capture local degradation characteristics caused by haze, such as blur diffusion and reduced contrast. Especially in areas where foreground objects or structures meet, the local detail estimation network effectively improves the spatial resolution and structural continuity of the transmittance map, avoiding the gradient mutation problem of transmittance estimation under blurred edges, laying the foundation for accurate physical modeling.

[0062] Finally, based on a classic atmospheric scattering model, the predicted physical parameters are integrated to reconstruct an indirect dehazed image based on the physical model. This image not only has good physical interpretability, avoiding "black-box" processing, but also complements the direct clear image generated by the content recovery branch, participating in subsequent comparative supervision and consistency constraints, effectively improving the overall robustness, generalization ability, and dehazing accuracy of the network.

[0063] In the design of feature contrast representation, a feature contrast learning mechanism was introduced in this phase to effectively decouple clear content features from haze blur features in the image. This mechanism decouples and models the encoded features in the intermediate layers between the content recovery branch and the parameter estimation branch, constructing a content feature stream representing the clear structural information of the image and a blur feature stream representing haze interference information. Based on this, a feature contrast loss function was designed based on the prior assumption that similar features tend to aggregate, while dissimilar features tend to repel. This loss function aims to aggregate semantically identical features in the latent feature space, bringing homologous information closer together; at the same time, it effectively separates features originating from different semantics (content and haze), enhancing their mutual exclusivity. Through this mechanism, the network adaptively learns the structural relationships between similar and different features during training, thereby strengthening the feature discrimination ability between the two branches and improving the overall feature decoupling capability of the network, laying a more stable semantic foundation for subsequent image dehazing and reconstruction tasks.

[0064] In the design of the reverse hazing architecture, a reverse hazing path based on an inverse atmospheric scattering model is proposed to establish an unsupervised cycle consistency mechanism. This architecture takes as input the direct dehazed image output by the content recovery branch and the atmospheric light parameters and transmittance map predicted by the parameter estimation branch. Combined with the classical atmospheric scattering physics model, it performs a reverse propagation operation to generate a simulated hazed image. Specifically, this module utilizes an inverse modeling approach to rehaze the dehazed image, giving it appearance characteristics similar to the original hazy image. Subsequently, the generated hazed image is aligned and constrained in pixel space with the original input image, forming a closed-loop structure of "hazing → dehazing → rehazing." This closed-loop mechanism effectively introduces self-supervised learning signals, enabling optimized training of the dehazing network without relying on clear image pairs. Furthermore, by enforcing the reversible consistency between the dehazing and hazing processes, the model's adaptability and generalization performance to complex haze variations in real scenes are significantly improved, providing more robust physical priors for the unsupervised image dehazing task.

[0065] In the design of pixel contrast representation, a pixel-level contrastive learning strategy was designed to further enhance the network's discriminative ability and generalization performance in the image pixel space domain. This strategy uses four types of image samples generated by the network, namely direct dehazed images, indirect dehazed images, reversely hazed images, and original haze images, as the basic sample set. It constructs multiple sets of triplet contrast relationships and uses the semantic similarity and difference features in the pixel space to guide the network to perform effective supervised learning. Specifically, this strategy constructs two sets of complementary contrastive learning paths:

[0066] In the first set of contrastive learning paths, the network uses the directly dehazed image as the anchor sample, the indirectly dehazed image reconstructed based on the physical model as the positive sample, and the original hazy image as the negative sample. This design leverages the contrast between the semantic and structural features of the images before and after dehazing to construct a "clear-interference" supervision framework. Its core goal is to guide the network in feature space: closing the gap between the directly dehazed image and the indirectly dehazed image, strengthening the convergence of the two paths in terms of clarity, structural fidelity, and semantic consistency; and furthering the gap between the two and the original image containing haze interference, thereby suppressing feature interference such as blur, low contrast, and color cast caused by residual haze. Through this construction of positive and negative samples, the network can effectively learn the distinguishing features between haze and non-haze during training, further improving its ability to discriminate and cluster "clear" image features.

[0067] In the second set of comparison paths, the network uses the reverse fogged image as the anchor sample, the original haze image as the positive sample, and the directly dehazed image as the negative sample. This path, starting from the perspective of "inverse process modeling," focuses on strengthening the similarity in feature space between the fogged image and the real haze image, thereby improving the realism and restoration ability of the fogged simulation. Its core mechanism is to: shorten the distance between the reverse fogged image and the real haze image to ensure sufficient simulation ability when modeling the physical degradation path; and further distance the fogged image from the clear image to strengthen its "degraded" characteristics, thereby enhancing the distinguishability between the clear and hazy states in the feature dimension. This path effectively provides symmetry constraints for the "dehazing-hazing" closed-loop learning process, helping to enhance the consistency and credibility of the two-way process.

[0068] By forming a collaborative training mechanism with the first set of paths (clear aggregation and haze removal), the network achieves: clear separation of foggy and clear images in feature space; improved consistency and reversibility of outputs during dehazing and fogging; and effective optimization of dehazing quality without relying on paired supervised labels. In summary, this dual-path pixel-level comparison mechanism not only achieves aggregation and separation of clear and hazy images in semantic space, but also significantly improves the network's discriminative, structural restoration, and unsupervised learning capabilities, providing a more universal and interpretable training paradigm for high-quality image dehazing tasks.

[0069] Based on the above network architecture and module design, in order to fully verify the effectiveness and advancement of the proposed unsupervised image dehazing network in practical applications, this paper conducted an experimental evaluation of the designed model and carried out multiple sets of comparative experiments with current mainstream image dehazing methods. The training, testing and evaluation process of the proposed dehazing network is shown in the attached figure. Figure 2 As shown. Specifically, 1,000 haze images were collected in real natural foggy scenes as the data basis, and divided into training set, validation set and test set in the ratio of 8:1:1, which were used for model training, parameter tuning and final performance evaluation respectively. In the training phase, 800 images were used to train the constructed unsupervised dehazing network end-to-end; in the validation phase, 100 images were selected for performance monitoring and model selection in the intermediate process; in the test phase, the remaining 100 images were used to independently evaluate the generalization ability of the model and the actual dehazing effect. In order to ensure the fairness and authority of the experimental comparison, three representative dehazing methods were selected as comparison objects: (1) the traditional atmospheric physics model method based on image prior, dark channel prior method DCP; (2) the typical supervised learning method GCA-Net; (3) the unsupervised dehazing method Cycle-Dehaze proposed in recent years. Qualitative comparison indicators include but are not limited to image clarity, detail retention ability, natural color restoration and overall visual quality. As shown in the attached figure. Figure 3 As shown in the results, the proposed unsupervised image dehazing network can achieve high-quality image restoration without the need for real clear images as supervisory signals. The generated dehazed images are significantly superior to the above-mentioned existing methods in multiple key dimensions such as structural edge preservation, texture reconstruction, and color consistency. This method shows stronger robustness and adaptability. The above experimental verification results fully demonstrate that the proposed dual-branch decoupling modeling architecture, feature contrast enhancement mechanism, and physical consistency-based loop optimization strategy are highly effective and advanced in image dehazing tasks. They not only improve the quality of image restoration under unsupervised conditions, but also provide a practical technical path for image enhancement and dehazing application tasks in real scenes.

[0070] To fully verify the effectiveness, robustness, and advancement of the unsupervised image dehazing network proposed in this invention in practical application scenarios, a series of systematic experimental evaluations were carried out based on the design of the above-mentioned network overall architecture and key functional modules. In addition, multiple groups of comparative experiments were conducted in combination with current mainstream image dehazing methods, aiming to objectively evaluate the performance of the network model of this invention from a qualitative dimension.

[0071] As attached Figure 2As shown, the overall evaluation process for the dehazing network employed in this invention, namely training-validation-testing, consists of five phases: data collection, data partitioning, model training, validation and parameter tuning, and independent testing. Specifically, experimental data was collected from foggy scenes in real natural environments, yielding a total of 1,000 haze image samples covering a variety of weather conditions, shooting angles, and terrain types to ensure data diversity and comprehensiveness for generalization performance evaluation. This data was proportionally divided into a training set (800 images), a validation set (100 images), and a test set (100 images), used for end-to-end training of the network model, hyperparameter tuning, and independent evaluation of the final model performance. The training phase employed an unsupervised learning strategy, jointly optimizing the various submodules of the network architecture to achieve high-quality feature extraction and image restoration capabilities. The validation phase monitored the network's convergence and generalization capabilities in real time and facilitated the selection of optimal model parameters. The testing phase tested the model's performance on a sample set that had never been used in training, ensuring the independence and reliability of the evaluation results.

[0072] To ensure the fairness and authority of the comparative experiment, three representative image dehazing methods were selected as reference standards: (1) Dark Channel Prior (DCP), an image prior algorithm based on atmospheric physics models, which represents the traditional prior modeling dehazing method; (2) GCA-Net, a typical end-to-end supervised learning network, which embodies the outstanding performance of the supervision scheme in the current deep learning method; and (3) Cycle-Dehaze, an unsupervised dehazing network that has emerged in recent years, which is used for horizontal comparison with the unsupervised model proposed in this paper under the same training conditions.

[0073] In terms of performance evaluation indicators, a comprehensive comparison was conducted using a number of qualitative evaluation criteria, including image clarity, texture retention, natural color reproduction, structural edge fidelity, and overall visual quality. The results are shown in the attached figure. Figure 3 As shown in the figure, the proposed unsupervised image dehazing network achieves high-quality image restoration without the need to provide real clear images as supervision signals, and performs superiorly in multiple key visual dimensions, especially in edge structure preservation, texture detail reconstruction and color consistency, which is significantly better than the comparison method, fully demonstrating the robustness and adaptability of this method in complex haze scenes.

[0074] Experimental results further demonstrate the high effectiveness and advancement of the proposed dual-branch decoupled modeling architecture, contrast-enhanced feature learning mechanism, and physically consistent closed-loop optimization strategy in real-world image dehazing tasks. Compared to existing unsupervised approaches, the network designed in this paper achieves more natural and realistic image restoration without relying on paired clear image data. This provides a highly feasible and adaptable technical path for implementing low-cost, high-quality image enhancement and dehazing systems in future real-world scenarios.

[0075] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bidirectional decoupled translation unsupervised dehazing network with feature and pixel contrast representation, characterized by: The following steps are involved: Step 1: Design of an overall unsupervised dehazing framework. This framework considers the image dehazing task as a feature decoupling process and constructs a dual-branch decoupling structure to separately learn the different feature distributions of haze information and clear content. The framework consists of a content recovery branch and a parameter estimation branch. The former focuses on restoring image details and structure, while the latter is used to estimate parameters related to the atmospheric scattering model. To improve the network's ability to distinguish haze and clear components in the latent feature space, a feature contrast learning mechanism is introduced to enhance the feature separation effect between branches. Simultaneously, by combining the physical modeling of forward dehazing and reverse hazing, a closed-loop structure is constructed to implement cycle consistency constraints, providing effective supervision signals under unsupervised conditions. The consistency constraints in the image space are further strengthened through a pixel-by-pixel comparison mechanism, thereby improving the network's discriminability and generalization capabilities. Step 2: Design a content restoration branch. By building a clear content feature mining module, the clear content information of the input haze image is extracted to obtain a directly dehazed image. Step 3: Parameter estimation branch design: By constructing modules for parameter estimation of the atmospheric scattering coefficient and transmittance respectively, the indirect dehazed image is restored based on the classic atmospheric scattering model. Step 4 : Feature contrast representation design, which decouples and models clear content features and haze features in the latent space, taking into account the inherent self-similarity in each information flow and the mutual exclusivity between different feature components; Step 5: Design a reverse fogging structure to perform reverse fogging on the direct dehazing result. Use the atmospheric scattering coefficient and transmittance predicted by the parameter estimation branch to reversely fog the direct dehazing result to generate a reverse fogged image. This image is then subjected to cycle consistency constraints with the original haze input image. Step 6: Design pixel contrast representation. Apply pixel-level contrast representation constraints to four types of images: direct dehazed images, indirect dehazed images, reversely hazed images, and original haze images, to enhance the expressiveness of the final dehazed results.

2. As described in step 1, the overall unsupervised dehazing framework is characterized by: treating dehazing as a feature decoupling task, that is, separating content-related information from clean factors, and separating haze-related information from fuzzy factors, realizing feature domain differentiation processing to obtain clear images, and providing a dehazing network structure paradigm that includes content recovery branch, parameter estimation branch, feature contrast representation, inverse hazing structure and pixel contrast representation, realizing end-to-end image detail restoration and physical modeling dehazing, and through multi-branch collaboration and dual output mechanism, generating direct dehazed images with high detail restoration capabilities and indirect dehazed images with physical consistency, thereby improving the network's dehazing performance and adaptability under unsupervised conditions.

3. As described in step 2, the content recovery branch is characterized by: a complete set of clear content feature mining mechanisms including a multi-level convolutional encoder, a residual feature extractor, a semantic enhancement unit, a frequency enhancement module, and an adaptive color correction module. The multi-level encoder is responsible for extracting texture, edge, and semantic features at different scales. The residual mechanism ensures the stable transmission of deep information. The semantic enhancement unit focuses on key areas through the attention mechanism. The frequency enhancement module strengthens the expression of high-frequency details. The adaptive color correction module dynamically adjusts the color tone to compensate for the color shift caused by haze.

4. As described in step 3, the parameter estimation branch is characterized by: estimating the atmospheric scattering coefficient and transmittance by exploring features related to haze, adopting a feature extraction paradigm with an inverted triangle structure, treating haze as a superposition of haze of different sizes and concentrations based on its uneven distribution, and exploring multi-scale haze-related features; designing a global context perception module, which effectively models the global brightness distribution and scene illumination changes in the image through multi-scale pooling or global attention mechanisms, thereby improving the stability and consistency of atmospheric light value prediction; designing a local detail estimation module, which enhances the transmittance modeling capability of local areas through a convolutional structure, focusing on the changes in degradation features in edge and texture areas, and improving the spatial accuracy of the transmittance map.

5. As described in step 4, the characteristics of the feature contrast representation design are: introducing a feature contrast learning mechanism, decoupling and modeling the intermediate encoding features of the content recovery branch and the parameter estimation branch, extracting the content feature stream and the haze blur feature stream of the image respectively, and designing a feature contrast loss function based on the contrast prior of aggregating similar features and separating heterogeneous features to achieve effective distinction between content and haze information in the feature space.

6. As described in step 5, the reverse hazing architecture is characterized by: based on the direct dehazing results obtained by the content recovery branch and the atmospheric scattering coefficient and transmittance predicted by the parameter estimation branch, the hazy image is reversely generated using an inverse atmospheric scattering model, forming a closed-loop structure of "hazing → dehazing → rehazing". This structure introduces a self-supervised cycle consistency signal and strengthens the reversibility constraints of the hazing and dehazing processes.

7. As described in step 6, the characteristics of the pixel contrast representation design are: a pixel-level contrast learning strategy is designed to construct two sets of triple contrast relationships based on the direct dehazed image, the indirect dehazed image, the reverse hazed image, and the original haze image, thereby strengthening the network's ability to distinguish image clarity in pixel space, and achieving effective separation of the feature space of clear images and haze images by setting complementary contrast paths.

Images