A non-paired image defogging and smoke removing method based on brightness prior and attention feature decoupling

Abstract

The application discloses a non-paired image defogging and smoke removing method based on luminance prior and attention feature decoupling, which comprises the following steps: obtaining a degraded image and extracting RGB three channels and luminance Y component, and splicing into a four-channel input tensor; adopting a double-branch parallel encoder to independently encode the RGB color feature and the luminance feature; using the luminance feature to generate a soft mask attention weight matrix through edge extraction and Sigmoid mapping at the feature intersection, and performing pixel-by-pixel modulation on the RGB feature to filter out highlights and heavy fog artifacts, and obtaining a decoupling feature tensor through residual fusion; and reconstructing a clear image through a PAN neck network and a mapper. The network is trained and constrained through a joint loss function containing a contrast estimation loss and a first-order spatial differential gradient edge preservation loss. Under the condition of not depending on paired training data, the application effectively restores the real color and sharp edges of the image, and can be widely deployed in embedded vision systems such as fire-fighting robots and industrial monitoring.

Description

Technical Field

[0001] This invention discloses a method for dehazing and removing smoke from unpaired images based on brightness priors and attention feature decoupling in the fields of computer vision, digital image processing and deep learning architecture design. Background Technology

[0002] Visual perception is the core information source for front-end intelligent applications such as autonomous mobile robots, explosion-proof video monitoring, and autonomous drone navigation to achieve 3D scene reconstruction and target guidance matching. In indoor fire rescue and high-risk industrial operation scenarios, mobile fire rescue robots rely heavily on their stereo vision-based guidance systems and real-time target detection frameworks to accurately locate dangerous fire sources and autonomously avoid irregular obstacles. However, due to the absorption and scattering effects of dense smoke from the incomplete combustion of chemicals in the air and a large number of suspended particles on spatial light, according to atmospheric scattering physics models, this optical scattering severely degrades the visibility and color contrast of images captured by cameras, obscuring high-frequency textures, geometric edges, and basic topological relationships of the scene. This not only leads to severe parallax calculation bias in binocular stereo vision matching algorithms but also causes a precipitous degradation in the performance of subsequent advanced semantic perception tasks (such as multi-class target recognition). Therefore, stably restoring high-fidelity physical scene structures from a single degraded perception image is a core requirement in the field of intelligent perception preprocessing.

[0003] Current technical approaches for image restoration in this category are mainly divided into methods based on physical statistical laws and learning paradigms based on deep learning. Traditional physical-driven models (such as dark channel prior DCP, color attenuation prior, etc.) rely on ideal natural image statistical prior laws to estimate atmospheric scattering model parameters. However, these approaches are poorly adaptable to complex and variable indoor lighting environments. Once a large area of ​​highly reflective metal, white equipment, or direct strong artificial light sources appears in the image, the physical prior assumptions will locally fail, leading to a serious deviation in the transmittance estimation, which in turn causes large-scale color distortion and shadow residue in the restored scene.

[0004] In recent years, supervised image restoration networks based on convolutional neural networks and visual Transformers have achieved significant technological breakthroughs. These models utilize end-to-end fitting mechanisms to fit large-scale paired synthetic datasets. However, most synthetic algorithms are based on the assumption of uniform atmospheric scattering, which is fundamentally different from the non-uniform dense smoke in real fire scenarios, characterized by drastic concentration changes and scattering from multi-band artificial light sources. Severe domain bias leads to significant performance degradation of supervised networks in practical deployments. While unsupervised unpaired translation paradigms (such as CycleGAN) alleviate the dependency bottleneck of paired datasets through bidirectional consistency cyclic constraints, their lack of explicit physical and boundary mathematical constraints in the feature space makes them prone to fabricating false high-frequency components in the dehazed image layer to meet numerical metrics for cyclic reconstruction during cross-domain mapping with unequal information entropy. The final output mask layer frequently exhibits severe detail breaks, deformations, and obvious geometric artifacts, failing to meet the stringent standards of modern industry for precise restoration.

[0005] In summary, existing mainstream single-image visual enhancement and unsupervised unpaired translation networks have the following significant adaptation shortcomings and technical drawbacks when dealing with realistic indoor non-uniform chemical smoke and extreme reflective scenes:

[0006] 1. Severe physical coupling in feature space: In the conventional RGB color space, the inherent spectral reflectance of the photographed object and the additive illumination degradation information introduced by atmospheric scattering exhibit a nonlinear high coupling. Existing single-branch generator architectures inevitably mix these two fundamentally different types of features together during downsampling and channel feature extraction, making it difficult to separate the lighting style during output reconstruction, resulting in severe color shift.

[0007] 2. Lack of explicit physical boundary constraints: Due to the severe asymmetry in information entropy between the clear target domain and the dense smoke degradation domain under unpaired conditions, existing one-sided or two-sided cyclic mapping algorithms lack explicit physical model mechanisms as boundary guidance. The models exhibit poor stability during adversarial training and distribution fitting, and the generated masks often suffer from edge defects and topological distortions.

[0008] 3. Limitations of Smoothing in Contrastive Learning: Represented by current mainstream implicit mutual information contrastive learning networks, when facing industrial disaster scenarios with highly homogenized backgrounds due to dense smoke coverage, the lack of strong boundary supervision at the micro-hierarchical level makes the contrastive loss prone to false negative penalties for low-contrast edges when repulsing feature blocks at different spatial locations. The model tends to sacrifice high-frequency details to reduce overall loss, resulting in blurred or even lost physical contours of highly reflective metal equipment and fine pipelines.

[0009] The aforementioned technical bottlenecks severely restrict the deployment effectiveness and engineering practical value of defogging and smoke removal algorithms in real-world harsh scenarios such as fire emergency rescue and industrial intelligent monitoring. Summary of the Invention

[0010] In view of this, the present invention discloses a non-paired image dehazing and smoke removal method based on the decoupling of brightness prior and attention features. By explicitly extracting the physical brightness orthogonal component as the environmental prior, designing a parallel dual-branch encoder to achieve hidden layer separation of color and shadow, and constructing a first-order spatial differential gradient edge loss to strengthen geometric topological constraints, this method solves the three core problems of color distortion, structural artifacts, and edge blurring without relying on pixel-level paired data. This provides a high-fidelity and robust preprocessing solution for intelligent visual perception under extreme conditions.

[0011] The technical solution provided by this invention is specifically as follows: This invention provides a method for dehazing and desmoking unpaired images based on the decoupling of brightness prior and attention features, comprising the following steps:

[0012] S1: Preprocessing stage: Acquire a single degraded image with fog or smoke, and perform preprocessing on the degraded image to convert it into a multi-channel input perceptual tensor;

[0013] S2: The multi-channel input sensing tensor is fed into a dual-branch parallel encoder, wherein the RGB encoding branch downsamples and encodes the RGB three-channel features to extract the inherent chromatographic distribution and low-frequency environmental components, and the brightness encoding branch downsamples and encodes the brightness Y component to extract the geometric density gradient and macroscopic physical contour boundary of suspended particles.

[0014] S3: At the intersection of the color feature map output by the RGB encoding branch and the luminance feature map output by the luminance encoding branch, a soft mask attention weight matrix is ​​generated using the luminance feature map output by the luminance encoding branch.

[0015] The color feature map output by the RGB encoding branch is modulated by pixel-wise bitwise multiplication using the soft mask attention weight matrix to filter out low-frequency redundant highlights and fog artifacts in the color features; the modulated color features are then deeply fused with the luminance feature map output by the luminance encoding branch through residual connection.

[0016] S4: The multi-scale composite decoupled feature tensor is transmitted to the end mapper through the PAN neck network for image reconstruction, and a clear restored image without fog or smoke is output.

[0017] The operations in steps S1 to S4 are performed by a brightness-aware attention decoupling network LAD-Net, which performs unpaired training constraint optimization through a joint loss function that includes a metric space contrast estimation loss based on a four-channel multi-scale latent feature space and an edge-preserving loss based on a Y-channel high-frequency first-order spatial calculus gradient.

[0018] Preferably, the preprocessing includes the following steps:

[0019] S11: Extract the RGB three-channel color information of the degraded image;

[0020] S12: The Y component of single-channel brightness is calculated according to the weighted brightness transformation formula;

[0021] ;

[0022] In the formula, This represents the single-channel luminance component calculated at pixel position x; , B(x) and B(x) represent the red, green, and blue channel pixel values ​​at pixel position x in the degraded image, respectively; the coefficients 0.299, 0.587, and 0.114 are the luminance weighting coefficients defined by the BT.601 standard.

[0023] S13: The RGB three channels and the luminance Y component are horizontally spliced ​​in the channel dimension to construct a multi-channel input perception tensor with a depth of 4.

[0024] Preferably, the generation of a soft mask attention weight matrix using the luminance feature map output from the luminance coding branch includes:

[0025] Edge extraction is performed on the input brightness feature map; Gaussian blur discretization operator is used to denoise and smooth the image.

[0026] ;

[0027] In the formula, Represents a two-dimensional Gaussian kernel function; The standard deviation of the Gaussian kernel is used to control the smoothing level; x and y represent the horizontal and vertical coordinate offsets of the pixel, respectively. This represents the image after Gaussian filtering and smoothing. Represents the input brightness feature map; This represents a two-dimensional discrete convolution operation;

[0028] By performing two-dimensional discrete convolution differentiation in the horizontal and vertical directions using the parameterless Sobel difference operator, a robust edge map and gradient magnitude that are resistant to noise are obtained:

[0029] ;

[0030] In the formula, , These represent the gradient component maps obtained after convolution using the Sobel horizontal and vertical difference operators, respectively. , These represent the 3×3 discrete difference convolution kernels in the horizontal and vertical directions of Sobel, respectively. This represents the brightness feature map after Gaussian smoothing. Represents the gradient magnitude plot;

[0031] The gradient magnitude map is mapped through a learnable convolutional layer and then processed by a sigmoid activation function to map it to a probability range of 0 to 1, thus constructing a soft mask attention weight matrix with clear spatial orientation:

[0032] .

[0033] In the formula, This represents the generated soft mask attention weight matrix, whose element values ​​are mapped to the interval [0,1]. This represents the luminance feature map output by the luminance encoding branch; , δ(·) represents the learnable convolutional layer weight parameters; δ(·) represents the ReLU nonlinear activation function; σ(·) represents the Sigmoid activation function. This represents the convolution operation.

[0034] Preferably, the deep fusion expression is:

[0035] .

[0036] In the formula, F fused M represents the multi-scale composite decoupled feature tensor output after deep fusion; attn F represents the soft mask attention weight matrix; rgb ⊗ represents the color feature map output by the RGB encoding branch; ⊗ represents element-wise product; ⊕ represents element-wise summation.

[0037] Preferably, the metric space contrastive estimation loss based on the four-channel multi-scale latent feature space is:

[0038] ;

[0039] ;

[0040] In the formula, This represents the contrastive estimation loss value for a single feature block; Represents the query feature vector; This represents the feature vector of a positive sample, located in the same spatial position as v; v n- τ represents the feature vector of the nth negative sample, which is located in a different spatial position from v; τ represents the temperature hyperparameter, used to control the concentration of the feature space distribution; N represents the total number of negative samples; · represents the vector dot product operation;

[0041] This represents the total loss of multi-scale patch-level contrast estimation; Indicates the input degraded image Calculate the expected value; L represents the total number of multi-scale feature layers; Indicates the first The feature layer index; s represents the s-th spatial position in the feature map of the l-th layer; This represents the total number of spatial locations in the feature map of layer l.

[0042] The edge-preserving loss based on Y-channel high-frequency first-order spatial differential gradient:

[0043] ;

[0044] In the formula, This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; , These represent the first-order spatial difference directional derivative discretization operators in the horizontal and vertical directions, respectively; This represents the luminance Y-channel component of the dehazed and smoke-restored image output by the generator; Represents the brightness Y-channel component of the sharp reference image in the target domain; ||·||1 represents the L1 norm;

[0045] Combining least-squares generative adversarial loss, the global optimization objective total loss function model of LAD-Net is expressed as:

[0046] .

[0047] In the formula, This represents the global optimization objective total loss function of the LAD-Net network; This represents the least squares generative adversarial loss (LSGAN). This represents the loss for multi-scale patch-level contrastive estimation; This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; This represents the balancing weight hyperparameter for the comparative estimation loss; This represents the balance weight hyperparameters of the edge-preserving loss.

[0048] This invention also provides an unpaired image dehazing and smoke removal system based on brightness prior and attention feature decoupling, characterized in that the above method includes:

[0049] The image acquisition module is used to acquire a single degraded image with fog or smoke.

[0050] An embedded processing module is deployed with a brightness-aware attention decoupling network. The network is loaded into the embedded processing module after being optimized by unpaired training constraints using the joint loss function described in claim 1. The network includes a channel stitching unit, a dual-branch parallel encoder, a gated spatial attention fusion unit, a neck network, and a mapper, which are used to perform dehazing and smoke-removing enhancement processing on the degraded image and output a clear restored image.

[0051] The display and transmission module is used to display the clear restored image in real time and transmit the clear restored image to the downstream target detection or stereo vision matching module.

[0052] This invention provides an unpaired image dehazing and smoke removal method based on the decoupling of brightness priors and attention features. It starts from the data perception layer, explicitly extracting orthogonal components of physical brightness as environmental priors injected into the network. A novel parallel dual-branch encoder architecture and spatially gated attention mechanism are employed to achieve the separation and dynamic modulation of inherent color representation and degraded lighting styles in the hidden layer. Simultaneously, this invention constructs a joint loss function based on implicit content contrast and high-frequency first-order spatial differential gradient edge constraints, providing strong geometric topological boundary guidance while completely eliminating interference from absolute brightness mean attenuation. This invention not only effectively solves the problem of unsupervised image restoration in the absence of pixel-paired data, restoring the true color and sharp physical edges of objects, but also undergoes rigorous validation using the published URSD fire smoke dataset, demonstrating that the proposed network can serve as a high-performance visual preprocessing module, significantly improving the spatial perception accuracy of downstream target detection models such as fire robots under extreme and harsh conditions.

[0053] Compared with the prior art, the present invention has the following significant advantages:

[0054] 1. Completely overcomes the strong dependence of traditional dehazing and smoke removal networks on pixel-level fully aligned paired datasets, endowing the model with high-fidelity image restoration capabilities in completely unsupervised, unpaired real disaster and special industrial scenarios.

[0055] 2. It specifically addresses the defects such as primary color shift and highlight overexposure caused by non-uniform dense smoke obstruction and strong reflection, realizes inherent color reconstruction, and reasonably protects the physical continuity of scene ambient lighting.

[0056] 3. It breaks the over-smoothing defect of the contrastive learning paradigm in a homogenized background, and uses differential loss to make the dehazed micro-textures and equipment components present clear and three-dimensional macro-structural edges.

[0057] 4. It possesses excellent software and hardware adaptability and multi-target recognition scalability, and can serve as a high-performance intelligent perception preprocessing module, significantly improving the spatial positioning accuracy of downstream target detection algorithms under harsh and complex industrial conditions.

[0058] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the disclosure of the present invention. Attached Figure Description

[0059] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This image shows a comparison of the visual effects of the LAD-Net method proposed in this invention and the traditional unsupervised unpaired translation model in foggy and strong light source scenes.

[0062] Figure 2 The diagram shows the overall architecture topology of the neural network designed for this invention, based on luminance prior injection and a parallel dual-branch encoder.

[0063] Figure 3 This is a schematic diagram illustrating the internal structure and computational process of the high-dimensional feature map activation mask and the high-frequency Sobel operator edge extraction of the gated spatial attention mechanism provided by this invention.

[0064] Figure 4 This is a comparison chart showing the qualitative visual restoration performance characteristics of the method of the present invention (LAD-Net) with DCP, AOD-Net, CycleGAN, CUT, and UCL-Dehaze on the synthetic test set SOTS images and the real unpaired test set URSD images.

[0065] Figure 5 Visual comparison of ablation experiments as the core modules of this invention are gradually introduced. Detailed Implementation

[0066] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0067] To effectively address the technical bottlenecks of scene geometric deformation and color cast caused by extreme non-uniform dense smoke occlusion and complex light source scattering coupling in the absence of absolutely paired reference images, this implementation scheme provides an unpaired image dehazing and smoke removal method based on brightness prior and attention feature decoupling. The core network architecture topology of LAD-Net is as follows: Figure 2 As shown, the system consists of a feedforward generator network that performs inverse cross-domain mapping, a discriminator network that supervises the distribution of latent feature space, and a multilayer perceptron that performs high-dimensional vector metric projection. The feedforward generator abandons the traditional single-branch design and consists of a C2fDarkNet backbone, an EgfFPN+PAN neck decoupling network, and a decoupling segmentation head.

[0068] The above methods specifically include:

[0069] S1: Preprocessing stage: Acquire a single degraded image with fog or smoke, and perform preprocessing on the degraded image to convert it into a multi-channel input perceptual tensor;

[0070] Based on the atmospheric scattering model, the image degradation process received by the camera is modeled using standard block physics equations:

[0071]

[0072] In the formula, I(x) represents the intensity of the degraded image received by the camera at pixel position x; J(x) represents the radiance of the scene of the clear, fog-free image to be recovered; t(x) represents the medium transmittance map, with a value range of [0,1]; A represents the global atmospheric light value; and x represents the two-dimensional pixel spatial coordinates in the image.

[0073] In the RGB color space, additive noise caused by scattering degrades lighting style and contaminates all three channels across the entire frequency range. To completely decouple physical features at the network front-end, this invention introduces an orthogonal decoupling strategy in the preprocessing stage, explicitly extracting the Y-channel feature map representing the absolute physical environment brightness variations in space.

[0074]

[0075] In the formula, This represents the single-channel luminance component calculated at pixel position x; , B(x) and B(x) represent the red, green, and blue channel pixel values ​​at pixel position x in the degraded image, respectively; the coefficients 0.299, 0.587, and 0.114 are the luminance weighting coefficients defined by the BT.601 standard.

[0076] In a physical optics environment, this Y-channel feature effectively eliminates interference from the intrinsic color spectral features of objects and accurately preserves the gradient density fluctuations of non-uniform suspended smoke. The network uses this single-channel brightness matrix as an explicit physical environment prior and concatenates it laterally with the conventional RGB pixel tensor in the channel dimension to construct a depth-4 depth input perception tensor, which serves as the signal source for a parallel multi-channel encoder.

[0077] This invention provides a physical brightness prior injection strategy based on spatial orthogonal decomposition of the data perception layer: a method for extracting the single-channel physical brightness Y component from the degraded image by weighting and eliminating spectral interference, and constructing a multi-channel perception feature vector by concatenating it with a regular RGB tensor, in order to cut off the coupling between color and light and shadow at the front end of the network input.

[0078] S2: The multi-channel input sensing tensor is fed into a dual-branch parallel encoder, wherein the RGB encoding branch downsamples and encodes the RGB three-channel features to extract the inherent chromatographic distribution and low-frequency environmental components, and the brightness encoding branch downsamples and encodes the brightness Y component to extract the geometric density gradient and macroscopic physical contour boundary of suspended particles.

[0079] S3: At the intersection of the color feature map output by the RGB encoding branch and the luminance feature map output by the luminance encoding branch, a soft mask attention weight matrix is ​​generated using the luminance feature map output by the luminance encoding branch.

[0080] The color feature map output by the RGB encoding branch is modulated by pixel-wise bitwise multiplication using the soft mask attention weight matrix to filter out low-frequency redundant highlights and fog artifacts in the color features; the modulated color features are then deeply fused with the luminance feature map output by the luminance encoding branch through residual connection.

[0081] To prevent feature aliasing between intrinsic color and degradation noise during downsampling convolution, the deep layers of the feedforward generator employ a dual-path parallel architecture. The RGB encoding branch is responsible for receiving the intrinsic color features of the three channels and extracting the intrinsic chromatographic distribution and low-frequency environmental components; while the Y luminance encoding branch focuses on receiving the prior feature plane and extracting the geometric density difference of suspended fog and the macroscopic physical contour boundary.

[0082] At the intersection of the two feature streams, this invention innovatively embeds a gated spatial attention decoupling and fusion network. The input layer feature map undergoes edge extraction. First, a Gaussian blur discrete operator is used to denoise and smooth the image.

[0083]

[0084] In the formula, Represents a two-dimensional Gaussian kernel function; The standard deviation of the Gaussian kernel is used to control the smoothing level; x and y represent the horizontal and vertical coordinate offsets of the pixel, respectively. This represents the image after Gaussian filtering and smoothing. Represents the input brightness feature map; This represents a two-dimensional discrete convolution operation;

[0085] Subsequently, the parameterless Sobel difference operator is used to perform two-dimensional discrete convolution differentiation in the horizontal and vertical directions to obtain a robust edge map and gradient magnitude that is resistant to noise. The specific formula is as follows:

[0086]

[0087] In the formula, , These represent the gradient component maps obtained after convolution using the Sobel horizontal and vertical difference operators, respectively. , These represent the 3×3 discrete difference convolution kernels in the horizontal and vertical directions of Sobel, respectively. This represents the brightness feature map after Gaussian smoothing. Represents the gradient magnitude plot;

[0088] The edge map is processed by a monotonically differentiable sigmoid activation function, mapping it to a probability range of 0 to 1, thus constructing a soft mask attention weight matrix with clear spatial orientation:

[0089]

[0090] In the formula, This represents the generated soft mask attention weight matrix, whose element values ​​are mapped to the interval [0,1]. This represents the luminance feature map output by the luminance encoding branch; , δ(·) represents the learnable convolutional layer weight parameters; δ(·) represents the ReLU nonlinear activation function; σ(·) represents the Sigmoid activation function. This represents the convolution operation.

[0091] The mask matrix and the feature stream of the RGB color branch are used to perform pixel-wise bitwise product modulation to filter out low-frequency redundant highlights and fog artifacts. Finally, it is merged with the Y luminance branch through residual transfer connection (element-wise summation). Its deep fusion expression is:

[0092]

[0093] In the formula, This represents the multi-scale composite decoupled feature tensor output after deep fusion; This represents the attention weight matrix for the soft mask. This represents the color feature map output by the RGB encoding branch; ⊕ indicates pixel-by-pixel bitwise multiplication; ⊕ indicates pixel-by-pixel bitwise addition.

[0094] This network aggregates features from shallow to deep, and, in conjunction with the subsequent PAN neck network, outputs multi-scale composite decoupled feature tensors to the end mapper.

[0095] The feedforward generator topology network structure provided by this invention features dual-branch attention feature decoupling and refusion: it protects the parallel use of independent color component extraction branches and luminance boundary extraction branches for cross-scale feature encoding, and embeds a gated spatial attention mechanism at the intersection of hidden layers (using luminance features to generate a spatially weighted attention mask to dynamically modulate and filter additive specular noise in the color stream), and finally performs deep fusion of the feedforward generator network topology relationship through multi-scale residual connections.

[0096] S4: The multi-scale composite decoupled feature tensor is transmitted to the end mapper through the PAN neck network for image reconstruction, and a clear restored image without fog or smoke is output.

[0097] The operations in steps S1 to S4 are performed by a brightness-aware attention decoupling network LAD-Net, which performs unpaired training constraint optimization through a joint loss function that includes a metric space contrast estimation loss based on a four-channel multi-scale latent feature space and an edge-preserving loss based on a Y-channel high-frequency first-order spatial calculus gradient.

[0098] Due to the lack of real pixel-level fully aligned reference images for strong supervision, this invention first constructs a metric space contrastive estimation loss based on a four-channel multi-scale latent feature space. By bringing the distance between the dehazed feature block and the original degraded feature block at the same location (positive sample pair) closer in the feature projection space, and strongly pushing apart feature blocks at different spatial locations (negative sample pairs), the mutual information in their spatial locations is maximized.

[0099]

[0100]

[0101] In the formula, This represents the contrastive estimation loss value for a single feature block; Represents the query feature vector; This represents the feature vector of a positive sample, located in the same spatial position as v; v n- τ represents the feature vector of the nth negative sample, which is located in a different spatial position from v; τ represents the temperature hyperparameter, used to control the concentration of the feature space distribution; N represents the total number of negative samples; · represents the vector dot product operation;

[0102] This represents the total loss of multi-scale patch-level contrast estimation; Indicates the input degraded image Calculate the expected value; L represents the total number of multi-scale feature layers; Indicates the first The feature layer index; s represents the s-th spatial position in the feature map of the l-th layer; This represents the total number of spatial locations in the feature map of layer l;

[0103] To completely overcome the "oversmoothing problem" caused by relying solely on latent space mutual information, this invention creatively designs a high-frequency first-order spatial calculus gradient edge-preserving loss (ECE Loss) based on the Y-channel. Since the partial derivative of the constant term is zero, this mechanism can completely filter out the bias caused by the decrease in absolute brightness value, focusing on strongly capturing and constraining geometrically abrupt edges in high-frequency scenes.

[0104]

[0105] In the formula, This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; , These represent the first-order spatial difference directional derivative discretization operators in the horizontal and vertical directions, respectively; This represents the luminance Y-channel component of the dehazed and smoke-restored image output by the generator; Represents the brightness Y-channel component of the sharp reference image in the target domain; ||·||1 represents the L1 norm;

[0106] Among them, the spatial first-order difference directional derivative discrete operator calculates the squared spatial distance between each erroneous pixel and the nearest real label edge for pixels that do not overlap between the two masks, and performs normalized weight allocation through the Softmax activation function to achieve highly sensitive calculus-based strong boundary protection.

[0107] Combining least-squares generative adversarial loss, the global optimization objective total loss function model of LAD-Net is expressed as:

[0108] .

[0109] In the formula, This represents the global optimization objective total loss function of the LAD-Net network; This represents the least-squares generative adversarial loss; This represents the loss for multi-scale patch-level contrastive estimation; This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; This represents the balancing weight hyperparameter for the comparative estimation loss; This represents the balance weight hyperparameters of the edge-preserving loss.

[0110] This invention utilizes a joint boundary constraint mechanism based on implicit spatial content contrast and first-order spatial differential gradient to protect the loss optimization constraint model in unpaired image enhancement. This model combines a contrast estimation loss function based on mutual information maximization of local spatial blocks in the image (used to constrain the consistency of multi-scale latent feature content) with a high-frequency gradient edge-preserving loss function calculated based on the first-order spatial difference directional derivatives of the Y channel (used to provide strong geometric boundary constraints to suppress smoothing artifacts).

[0111] This invention proposes an unpaired image enhancement algorithm based on brightness prior and attention decoupling, which can be seamlessly implemented and integrated into various types of special equipment image enhancement systems. As a typical embodiment, this invention is used to build a hardware platform for an image enhancement system targeting the front-end perception of a fire-fighting robot. Its basic structure includes a robotic arm, a mobile platform, an internal adaptive power supply module, a light source, fire-fighting equipment interaction tools, and a Linux embedded computing development board (Jetson Orin NX platform). During system operation, the camera module first captures bright, smoke-laden images of fire drills or accident scenes and transmits them in real-time to the embedded development board via a high-speed bus. The development board then uses the LAD-Net network deployed in memory to perform unpaired cross-domain smoke removal and restoration on the input tensor. The reconstructed high-definition, clear perception image is transmitted to the target detection module, and the restoration result is displayed in real-time. This system, combined with the light source, effectively resists physical interference from drastic fluctuations in external lighting, meeting the stable operation requirements under extreme and special working conditions.

[0112] To objectively and fairly evaluate the real-world reconstruction capability of LAD-Net, this invention utilizes Hikvision industrial cameras to collect and construct a real unpaired indoor fire smoke dataset (URSD) containing 2500 smoke degradation images and 2500 smoke-free reference images.

[0113] The experiment was conducted on a single NVIDIA RTX 3080 Ti graphics card platform equipped with 12GB of video memory, and was deployed and verified on a JETSON ORIN NX development board. The quantitative values ​​for the multi-algorithm quantitative comparison and ablation component stripping experiments are recorded in Tables 1, 2, 3, and 4, respectively.

[0114] Table 1. Comparison of multi-algorithm quantization results on the paired synthetic dataset SOTS-Outdoor

[0115]

[0116] Table 2. Comparison of multi-algorithm quantization results on the real unpaired fire smoke dataset URSD.

[0117]

[0118] Table 3. Progressive ablation quantitative analysis of network components on the synthetic dataset SOTS-Outdoor

[0119]

[0120] Table 4. Quantitative analysis of progressive ablation of network components on the unpaired real dataset URSD

[0121]

[0122] The comparative analysis of Tables 1 and 2 shows that supervised networks exhibit significant degradation in objective metrics when faced with complex fire-fighting non-uniform dense smoke scenarios, highlighting a serious bottleneck in cross-domain generalization. Traditional contrastive learning methods (such as UCL-Dehaze) suffer severe NIQE degradation to 8.3022 when faced with heavy, large-area smoke occlusion, demonstrating obvious feature mapping bias. In contrast, the LAD-Net proposed in this invention achieves superior and balanced quantitative feedback in both full-reference and no-reference blind evaluation systems, effectively removing non-uniform dense smoke occlusion while maintaining excellent image quality and naturalness restoration.

[0123] The component stripping and ablation experiments in Tables 3 and 4 further confirm that the original baseline (Model A) lacking brightness decoupling suffers from severe color distortion and highlight overexposure defects. Individual channel expansion (Model B) easily leads to deep coupling conflicts between color and brightness, resulting in smooth aliasing of image features. However, by fusing a bi-branch attention network with explicit spatial gradient derivative constraints (Model D), the network effectively achieves high-fidelity reconstruction of the intrinsic color spectrum, and makes the physical edges of distant macroscopic objects such as metal pipes and instrument gaps clear and sharp, fully demonstrating the engineering guiding value of each innovative strategy in complex unsupervised image restoration.

[0124] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

Claims

1. A method for dehazing and desmoking unpaired images based on brightness prior and attention feature decoupling, characterized in that, Includes the following steps: S1: Preprocessing stage: Acquire a single degraded image with fog or smoke, and perform preprocessing on the degraded image to convert it into a multi-channel input perceptual tensor; S2: The multi-channel input sensing tensor is fed into a dual-branch parallel encoder, wherein the RGB encoding branch downsamples and encodes the RGB three-channel features to extract the inherent chromatographic distribution and low-frequency environmental components, and the brightness encoding branch downsamples and encodes the brightness Y component to extract the geometric density gradient and macroscopic physical contour boundary of suspended particles. S3: At the intersection of the color feature map output by the RGB encoding branch and the luminance feature map output by the luminance encoding branch, a soft mask attention weight matrix is ​​generated using the luminance feature map output by the luminance encoding branch. The color feature map output by the RGB encoding branch is modulated by pixel-wise bitwise multiplication using the soft mask attention weight matrix to filter out low-frequency redundant highlights and fog artifacts in the color features. The modulated color features are deeply fused with the luminance feature map output by the luminance encoding branch through residual connection; S4: The multi-scale composite decoupled feature tensor is transmitted to the end mapper through the PAN neck network for image reconstruction, and a clear restored image without fog or smoke is output. The operations in steps S1 to S4 are performed by a brightness-aware attention decoupling network LAD-Net, which performs unpaired training constraint optimization through a joint loss function that includes a metric space contrast estimation loss based on a four-channel multi-scale latent feature space and an edge-preserving loss based on a Y-channel high-frequency first-order spatial calculus gradient.

2. The method for dehazing and desmoking unpaired images based on brightness prior and attention feature decoupling as described in claim 1, characterized in that, The preprocessing includes the following steps: S11: Extract the RGB three-channel color information of the degraded image; S12: Calculate the single-channel brightness according to the weighted brightness transformation formula. Quantity; ； In the formula, This represents the single-channel luminance component calculated at pixel position x; , B(x) and B(x) represent the red, green, and blue channel pixel values ​​at pixel position x in the degraded image, respectively; the coefficients 0.299, 0.587, and 0.114 are the luminance weighting coefficients defined by the BT.601 standard. S13: The RGB three channels and the luminance Y component are horizontally spliced ​​in the channel dimension to construct a multi-channel input perception tensor with a depth of 4.

3. The method for dehazing and desmoking unpaired images based on brightness prior and attention feature decoupling as described in claim 1, characterized in that, The soft mask attention weight matrix is ​​generated using the luminance feature map output from the luminance coding branch, including: Edge extraction is performed on the input brightness feature map; Gaussian blur discretization operator is used to denoise and smooth the image. ； In the formula, Represents a two-dimensional Gaussian kernel function; The standard deviation of the Gaussian kernel is used to control the smoothing level; x and y represent the horizontal and vertical coordinate offsets of the pixel, respectively. This represents the image after Gaussian filtering and smoothing. Represents the input brightness feature map; This represents a two-dimensional discrete convolution operation; By performing two-dimensional discrete convolution differentiation in the horizontal and vertical directions using the parameterless Sobel difference operator, a robust edge map and gradient magnitude that are resistant to noise are obtained: ； In the formula, , These represent the gradient component maps obtained after convolution using the Sobel horizontal and vertical difference operators, respectively. , These represent the 3×3 discrete difference convolution kernels in the horizontal and vertical directions of Sobel, respectively. This represents the brightness feature map after Gaussian smoothing. Represents the gradient magnitude plot; The gradient magnitude map is mapped through a learnable convolutional layer and then processed by a sigmoid activation function to map it to a probability range of 0 to 1, thus constructing a soft mask attention weight matrix with clear spatial orientation: ； In the formula, This represents the generated soft mask attention weight matrix, whose element values ​​are mapped to the interval [0,1]. This represents the luminance feature map output by the luminance encoding branch; , δ(·) represents the learnable convolutional layer weight parameters; δ(·) represents the ReLU nonlinear activation function; σ(·) represents the Sigmoid activation function. This represents the convolution operation.

4. The method for dehazing and desmoking unpaired images based on brightness prior and attention feature decoupling as described in claim 1, characterized in that, The deep fusion expression is: ； In the formula, This represents the multi-scale composite decoupled feature tensor output after deep fusion; This represents the attention weight matrix for the soft mask. This represents the color feature map output by the RGB encoding branch; ⊕ indicates pixel-by-pixel bitwise multiplication; ⊕ indicates pixel-by-pixel bitwise addition.

5. The method for dehazing and desmoking unpaired images based on brightness prior and attention feature decoupling as described in claim 1, characterized in that, The metric space contrastive estimation loss based on the four-channel multi-scale latent feature space: ； ； In the formula, This represents the contrastive estimation loss value for a single feature block; Represents the query feature vector; This represents the feature vector of a positive sample, located in the same spatial position as v; v n- τ represents the feature vector of the nth negative sample, which is located in a different spatial position from v; τ represents the temperature hyperparameter, used to control the concentration of the feature space distribution; N represents the total number of negative samples. • Represents the vector dot product operation; This represents the total loss of multi-scale patch-level contrast estimation; Indicates the input degraded image Calculate the expected value; L represents the total number of multi-scale feature layers; Indicates the first The feature layer index; s represents the s-th spatial position in the feature map of the l-th layer; This represents the total number of spatial locations in the feature map of layer l; The edge-preserving loss based on Y-channel high-frequency first-order spatial differential gradient: ； In the formula, This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; , These represent the first-order spatial difference directional derivative discretization operators in the horizontal and vertical directions, respectively; This represents the luminance Y-channel component of the dehazed and smoke-restored image output by the generator; Represents the brightness Y-channel component of the sharp reference image in the target domain; ||·||1 represents the L1 norm; Combining least-squares generative adversarial loss, the global optimization objective total loss function model of LAD-Net is expressed as: ； In the formula, This represents the global optimization objective total loss function of the LAD-Net network; This represents the least-squares generative adversarial loss; This represents the loss for multi-scale patch-level contrastive estimation; This represents the edge-preserving loss of the high-frequency first-order spatial differential gradient; This represents the balancing weight hyperparameter for the comparative estimation loss; This represents the balance weight hyperparameters of the edge-preserving loss.

6. A non-paired image dehazing and smoke removal system based on brightness prior and attention feature decoupling, characterized in that, The method according to any one of claims 1 to 5 comprises: The image acquisition module is used to acquire a single degraded image with fog or smoke. An embedded processing module is deployed with a brightness-aware attention decoupling network. The network is loaded into the embedded processing module after being optimized by unpaired training constraints using the joint loss function described in claim 1. The network includes a channel stitching unit, a dual-branch parallel encoder, a gated spatial attention fusion unit, a neck network, and a mapper, which are used to perform dehazing and smoke-removing enhancement processing on the degraded image and output a clear restored image. The display and transmission module is used to display the clear restored image in real time and transmit the clear restored image to the downstream target detection or stereo vision matching module.

Images