A Weather-Regional Degradation Co-guided Infrared Image Restoration Method under Multiple Weather Conditions

The infrared image restoration method guided by weather-region degradation solves the problem of infrared image restoration under multiple weather scenarios, achieving efficient and balanced image restoration results. It is applicable to fields such as autonomous driving, remote sensing monitoring, drone inspection, and security monitoring.

CN123118409APending Publication Date: 2026-11-17XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611220888.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing infrared image restoration technologies cannot effectively adapt to various weather degradations, resulting in reduced accuracy in target detection, segmentation, and recognition tasks. Furthermore, the models have poor versatility and high deployment costs, making it impossible to achieve unified restoration in complex outdoor scenarios.

Method used

By constructing a weather-region degradation collaboratively guided infrared image restoration method under multiple weather conditions, this method utilizes a weather semantic encoder and local cue priors, combined with proxy labels for the degree of regional degradation, to adaptively allocate restoration capabilities and achieve differentiated restoration of different regions within the image. A phased training strategy and multi-level cue adjustment are adopted to reduce annotation costs.

Benefits of technology

It achieves high-quality restoration of infrared images under multiple weather scenarios, alleviates over-restoration and under-restoration problems, maintains the continuity of target contours and thermal structures, improves the balance and robustness of restoration, and reduces cross-weather negative migration and expert call imbalance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118409A_ABST
    Figure CN123118409A_ABST
Patent Text Reader

Abstract

This invention relates to a weather-region degradation collaboratively guided infrared image restoration method under multiple weather conditions. It acquires infrared images degraded by rain, fog, snow, and cloud cover, and encodes them to obtain multi-scale features and a task view. High-frequency information is extracted and input into a weather semantic encoder to obtain a global weather prior. The decoder fuses the global weather prior and local features at each stage to generate adaptive cue modulation features. Based on clear-degraded image pairs, unannotated regional degradation proxy labels are constructed to supervise the prediction of regional degradation levels. Each decoding branch sets up a shared branch and a hierarchical weather expert pool, selecting the expert pool based on weather type and matching the corresponding complexity expert based on degradation level. Iterative decoding outputs the restored image by fusing multiple features, and a two-stage, multi-level training method is used to optimize the network. This invention achieves differentiated image restoration, weakens the coupling between thermal structure and weather degradation, reduces annotation costs, suppresses cross-weather negative transfer, and is adaptable to multi-weather infrared image restoration scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of infrared image processing technology, specifically relating to a method for infrared image restoration guided by weather-region degradation under multiple weather conditions. Background Technology

[0002] Thermal infrared imaging relies on the target's own thermal radiation, making it adaptable to complex lighting conditions such as nighttime, low light, and backlighting. It boasts strong imaging stability and is widely used in visual perception fields such as autonomous driving, remote sensing monitoring, drone inspection, and security surveillance. However, weather interference such as rain, fog, snow, and cloud cover can cause degradation issues in infrared images, including decreased thermal contrast, blurred target boundaries, local occlusion, and non-uniform structural distortion. This significantly reduces the accuracy of subsequent target detection, segmentation, and recognition tasks, limiting the practical performance of infrared imaging systems in complex multi-weather scenarios. Existing image restoration technologies are mostly designed separately for single weather degradations such as defogging, de-raining, and de-snowing. Traditional prior methods and single depth models can only adapt to preset weather scenarios, requiring the construction of independent models for different degradations. This results in poor versatility, high model redundancy, and high deployment costs, making it impossible to handle multiple weather degradations with a unified model and failing to meet the restoration needs of complex outdoor scenarios. To address the insufficient adaptability of single models, existing integrated image restoration technologies achieve unified restoration of multiple degradations through shared backbone networks, weather embedding, cue vectors, and hybrid experts. Currently, similar existing technologies can be mainly divided into two categories: one is the multi-weather integrated restoration method for visible light, which relies on global degradation type identification and matching repair strategies; the other is the traditional infrared image enhancement method, which mainly deals with inherent defects such as noise, blur, and low contrast. Only a few studies have been conducted on the restoration of single-weather infrared images, and both types of methods have obvious limitations.

[0003] However, existing technologies generally adopt a whole-image homogenization restoration strategy, relying solely on global weather categories to regulate the restoration process. They fail to perform hierarchical and refined modeling of the differences in degradation mechanisms under different weather conditions and the degree of non-uniform degradation within the image's internal space. Consequently, they cannot adapt to the differentiated degradation characteristics of infrared images with various weather conditions. Specific technical shortcomings are as follows: First, weather degradation is highly coupled with infrared thermal structure, making restoration prone to distortion. The outlines of weak targets and the low-contrast thermal response in infrared images are similar to the characteristics of weather interference. Existing technologies cannot accurately distinguish between effective thermal targets and degradation noise, easily suppressing the features of real weak targets, resulting in broken target edges, distorted thermal intensity, excessive smoothing of details, and loss of key sensing information.

[0004] Second, the lack of a coordinated regulation mechanism between weather type and regional degradation level results in poor restoration balance. Existing methods rely solely on global weather labels and uniform restoration parameters, which cannot distinguish degradation differences under different weather conditions, nor can they adapt to varying degrees of regional degradation within a single frame image. This easily leads to problems such as over-restoration in mild areas and residual restoration in severe areas, resulting in insufficient restoration robustness.

[0005] Third, the stability of regional degradation supervision and expert routing mechanisms is insufficient. Existing integrated models rely on implicit feature learning to complete degradation discrimination and expert selection, lacking explicit supervision of regional degradation levels. This results in high annotation costs, high training difficulty, and problems such as cross-weather negative transfer and imbalance in expert calling, leading to poor model generalization and multi-scenario repair stability.

[0006] Therefore, there is an urgent need for an infrared image restoration method that can adaptively allocate restoration capabilities to different degraded regions within the same image, reducing over-restoration and under-restoration, in order to solve the aforementioned technical problems. Summary of the Invention

[0007] This invention provides the following technical solution: a method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions, comprising the following steps: Step 1: Obtain the infrared image to be restored containing at least one weather degradation caused by rain, fog, snow, or cloud cover. Input the infrared image to be restored into the encoder to obtain multi-scale coded features and bottleneck features. Scaling the infrared image to be restored yields a task-level view.

[0008] Step 2: After concatenating the task-level view with its high-frequency response, input it into the weather semantic encoder to obtain the weather type logical quantity and the global weather semantic prior. The global weather semantic prior is used to limit the prompt selection and the selection of the weather-specific expert pool.

[0009] Step 3: At each stage of the decoder, local cue priors are generated based on the local recovery features of the current stage. The weather conditional cue distribution is calculated by combining the global weather semantic priors. Stage adaptive cue is sparsely selected from the cue library of the corresponding stage and fused with the features of the current stage to obtain cue modulation features. Deep cue modulates the global recovery direction, mid-level cue optimizes thermal contrast and main structure, and shallow cue refines edge texture.

[0010] Step 4: During the training phase, a proxy label for the degree of regional degradation is constructed based on the spatially aligned degraded infrared image and the clear infrared image. The proxy label is determined by the regional residual amplitude and residual gradient statistics, and conditional normalization is performed according to the weather type to eliminate statistical offsets between different weather categories.

[0011] Step 5: Based on the global semantic information and high-frequency degradation information of the encoder bottleneck features, combined with the global weather semantic prior prediction of the degradation degree score of the region, the prediction results are supervised by surrogate labels, so that the model can learn the spatial non-uniform degradation distribution without the need for manual labeling of the region degradation degree.

[0012] Step 6: Perform multi-source collaborative expert selection at each decoding stage: extract the general thermal structure across weather conditions through the shared recovery branch, select the target expert pool from multiple weather-specific expert pools through weather semantic priors, and then select experts from multiple complexity levels of experts in the target expert pool that match the current degree of regional degradation based on the degree of regional degradation.

[0013] Step 7: Fuse the cue modulation features, shared recovery features, and expert recovery features, and generate the features for the next decoding stage step by step through the feedforward enhancement module, and finally output the recovered infrared image.

[0014] Step 8: Adopt a phased training strategy: The first phase trains the weather semantics, phase prompts, and basic recovery network; the second phase introduces degradation degree estimation, type-degree hierarchical routing, degree consistency, and expert equilibrium constraints to obtain a stable and regionally adaptive multi-weather infrared image recovery model.

[0015] Preferably, in step 1, the task-level view is obtained from the infrared image to be recovered through bilinear interpolation.

[0016] Preferably, in step 2, the high-frequency response is obtained by performing a convolution operation on the task-level view using a Laplacian convolution kernel, and the concatenated expression is:

[0017] in, For task-level view, The symbol represents the Laplacian convolution kernel, and * represents the convolution operation.

[0018] Preferably, in step 3, the process of generating weather conditional prompt distribution is as follows: global average pooling of the features of the decoding stage is used to obtain the stage local prompt prior, the global weather semantic prior is concatenated and then a scaling factor is generated through a multilayer perceptron to calibrate the prompt selection logic quantity, a Top-1 sparse strategy is used to select a single stage adaptive prompt, and the feature is fused by a Transformer module and a 1×1 convolution.

[0019] Preferably, in step 4, the method for constructing the region degradation proxy label is as follows: dividing the image into independent regions and calculating the residual magnitude between the degraded region and the clear ground truth region. Horizontal gradient mean Vertical gradient mean Original degree of degradation ; Statistically calculate the average degree of degradation of similar weather samples Standard deviation The proxy label is obtained after Sigmoid normalization: ; in For gradient weights, The normalized temperature coefficient, is the numerical stability constant.

[0020] Preferably, in step 5, the degradation representation is obtained by: obtaining a semantic representation of the bottleneck features through global average pooling, and obtaining a frequency embedding through high-pass convolution, activation, pooling and multilayer perceptron, and adding the two to form a degradation representation; concatenating the degradation representation with the weather semantic prior, inputting it into the degradation degree prediction head, and outputting the regional degradation degree prediction value.

[0021] Preferably, in step 6, the specific process of multi-source collaborative expert selection includes: Weather semantic priors are used to perform sparse selection at the task level to determine a dedicated pool of weather experts. Set a degradation level center for each complexity level expert within each weather expert pool; Based on the matching bias between the degree of regional degradation and the center of degradation of each expert, combined with stage features and frequency embedding, the expert routing logic quantity is obtained, and the complexity routing weight is obtained through temperature scaling and sparse selection. A two-dimensional gating system is formed by task-level weights and degradation-level weights to control the output of expert branches.

[0022] Even better, experts at different complexity levels can form a gradient of recovery capability from light to heavy by adjusting the kernel size, residual stacking depth, and parameter scale.

[0023] Preferably, in step 8, the two-stage training loss function is set as follows: The total loss in the first phase is:

[0024] in For pixel reconstruction loss; Cross-entropy for weather categories; To compare and suggest recovery of losses, , These are the weighting coefficients. The total loss in the second phase is:

[0025] in, To monitor the degree of degradation and loss, For degradation of consistency loss, To balance the losses for the experts, , , These are the weighting coefficients.

[0026] Preferably, the encoder-decoder main network, the weather semantic encoder, the degradation degree prediction head, the multi-level prompt generation module, and the multi-level expert pool together constitute the infrared image restoration network; the network can simultaneously process four types of weather degradation: rain, fog, snow, and cloud cover, and output restored infrared images for target detection, segmentation, and recognition visual tasks.

[0027] The beneficial effects of this invention are: 1. This invention constructs a hierarchical control mechanism that determines the restoration direction based on weather type and the restoration intensity based on the degree of regional degradation. It can adaptively allocate the model restoration capability according to the degradation differences of different regions within a single frame image, and perform differentiated repair on different degraded regions, effectively alleviating the problems of over-restored and under-restored images, and improving the overall restoration balance of infrared images under multiple weather scenarios.

[0028] 2. This invention integrates weather semantics, stage local state, and multi-level prompt information, and can perform layered adjustments on the global repair direction, image thermal contrast, and target detail edges at different decoding depths of the network. This effectively alleviates the feature coupling problem between weather degradation interference and the true infrared thermal structure, reduces the risk of misjudging effective target features as noise and suppressing them, and better preserves the true target contours and thermal radiation details of the infrared image.

[0029] 3. This invention constructs weather-conditional proxy labels based on the residual amplitude and gradient statistics of paired regions, eliminating the need for extensive manual annotation. By combining shared branches, weather expert pools, and degradation level experts working together, it ensures the stability of the model's expert routing logic with low annotation costs, suppresses cross-weather negative migration and expert call imbalance, and still achieves excellent infrared image restoration quality in complex multi-weather degradation scenarios. Attached Figure Description

[0030] Fig. 1 This is a diagram of the overall framework of the restoration network for a weather-region degradation collaboratively guided infrared image restoration method under multiple weather conditions, as described in this invention. Fig. 2 This is a schematic diagram of the weather and semantic collaborative multi-level prompt generation module of the present invention; Fig. 3 This is a schematic diagram of the agent supervision area degradation severity estimation module of the present invention; Fig. 4 This is a schematic diagram of the multi-source collaborative expert selection strategy of the present invention; Fig. 5 This is a comparison chart of the qualitative results of the present invention and various integrated image restoration methods in the tasks of deraining rain, defogging, desnowing, and declouding. Detailed Implementation

[0031] The relevant technologies of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0032] like Figs. 1-5 As shown, the method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to this embodiment includes the following steps: Step 1) Obtain an infrared image to be restored containing at least one weather degradation caused by rain, fog, snow, or cloud cover. Input the infrared image to be restored into an encoder to obtain multi-scale coding features and bottleneck features. Simultaneously, scale the infrared image to be restored to obtain a task-level view. Step 2) Concatenate the task-level view with its high-frequency response and input it into the weather semantic encoder to obtain the weather type logical quantity and the global weather semantic prior. The global weather semantic prior is used to limit the selection of subsequent prompts and the selection of the weather-specific expert pool. Step 3) In multiple stages of the decoder, local cue priors are generated based on the local recovery features of the current stage, and the weather conditional cue distribution is calculated together with the global weather semantic prior. Stage adaptive cue is sparsely selected from the cue library of the corresponding stage, and the cue is fused with the features of the current stage to obtain cue modulation features. Step 4) During the training phase, a proxy label for the degree of regional degradation is constructed based on the spatially aligned degraded infrared image and the clear infrared image. The proxy label is determined by the regional residual amplitude and residual gradient statistics, and conditional normalization is performed according to the weather type to eliminate statistical offsets between different weather categories. Step 5) Based on the global semantic information and high-frequency degradation information of the encoder bottleneck features, combined with the global weather semantic prior prediction of the degradation degree score of the region, the prediction results are supervised by the proxy label, so that the model can learn the spatial non-uniform degradation distribution without the need for manual labeling of the region degradation degree. Step 6) Perform multi-source collaborative expert selection at each decoding stage: extract the general thermal structure across weather by sharing the recovery branch, select the target expert pool from multiple weather-specific expert pools by weather semantic priors, and then select experts from multiple complexity levels of experts in the target expert pool that match the current degree of regional degradation according to the degree of regional degradation. Step 7) The cue modulation features, shared recovery features, and expert recovery features are fused together, and the features for the next decoding stage are generated step by step through the feedforward enhancement module, and finally the recovered infrared image is output. Step 8) A phased training strategy is adopted: the first phase trains the weather semantics, phase prompts and basic recovery network; the second phase introduces degradation degree estimation, type-degree hierarchical routing, degree consistency and expert equilibrium constraints to obtain a stable and regionally adaptive multi-weather infrared image recovery model.

[0033] Specifically, in step 1), let the infrared image to be recovered be... The main recovery branch employs an encoder-decoder structure. The encoder outputs multi-scale features. and bottleneck characteristics Task-level view By bilinear interpolation get.

[0034] In step 2), to highlight edge attenuation, local occlusion, and structural disturbances caused by weather, the task-level view is stitched together with the high-frequency response:

[0035] in, The asterisk (*) represents the Laplacian convolution kernel, and the asterisk (*) represents the convolution operation. Weather semantic encoder. right Encoding is performed to obtain the weather type logical variable z, and the weather semantic prior is obtained through temperature scaling Softmax: where, For weather-related temperature parameters, Each dimension corresponds to a weather category such as rain, fog, snow, or cloud cover.

[0036] In step 3), in the first Each decoding stage, current stage features The stage representation is obtained through global average pooling. And map it to the prompt selection space:

[0037] Local hints prior Semantic Priors with Weather The data is then stitched together, and a positive scaling factor is generated using a multilayer perceptron. Finally, the logic for selecting the prompt is calibrated.

[0038] in, To ensure the scaling factor is a positive constant, let the first... The stage hint library is Hints were obtained using Top-1 sparse selection. :

[0039] The selected cues are scale-aligned and convolutionally mapped before being fused with the current stage features to obtain stage-adaptive cue modulation features:

[0040] Deep cues are used to define the global restoration direction, mid-level cues are used to restore thermal contrast and main structure, and shallow cues are used to refine edges, textures and local thermal response, thereby achieving a progressive decoupling from deep to shallow.

[0041] In step 4), the training image is divided into multiple regions. For any given region... Let the clear area of ​​spatial alignment be... Degraded areas are First, calculate the residual r and its average amplitude. :

[0042] Then calculate the gradient changes of the residuals in the horizontal and vertical directions:

[0043]

[0044] The intensity shift and structural perturbation are combined to form the original degradation level:

[0045] Since the same residual amplitude may correspond to different actual degradation levels in different weather conditions, the mean of the training samples belonging to weather category c is statistically analyzed. and standard deviation The degree of regional degradation caused by weather conditions is represented by the following proxy tags:

[0046] in, For gradient term weights, The severity normalization temperature, This is a numerical stability constant. This method eliminates the need for manual labeling of degradation areas and can emphasize the differences in relative degradation levels within the same weather category.

[0047] In step 5), bottleneck characteristics Semantic representation is obtained through global average pooling. Frequency embeddings are obtained through high-pass convolution, activation, pooling, and multilayer perceptron. The two together form the degenerative characteristics. :

[0048] Will Semantic Priors with Weather The algorithm concatenates the data, inputs a degradation prediction header, and outputs predicted degradation values ​​for the region. :

[0049] In step 6), the recovery block of each decoding stage includes a shared recovery branch and a hierarchical expert recovery branch. The shared branch normalizes and performs attention transformation on the cue modulation features to recover the weather-wide thermal structure.

[0050] Weather semantic priors are first implemented at the task level using sparse selection to determine a dedicated pool of weather experts.

[0051] Each weather expert pool contains multiple experts of varying complexity levels, and for the first... An expert set up a degradation degree center Regarding the degree of regional degradation Construct matching bias:

[0052] The expert routing logic quantity is obtained by combining stage features, frequency embedding, and degradation degree matching bias. The complexity routing weight is obtained by temperature scaling and sparse selection.

[0053] The task-level weights and degradation-level weights form a two-dimensional gating system, and the expert branch output is:

[0054]

[0055] in, This represents the t-th weather expert in the pool. Experts of varying complexity. Different levels of experts can be configured to provide varying recovery capabilities, from light to heavy, by adjusting kernel size, residual stacking depth, and parameter scale. Low-complexity experts are used for light-duty regions, while high-complexity experts are used for heavy-duty regions.

[0056] In step 7), the cue modulation features, shared branch features, and expert branch features are fused to obtain the output of the current stage:

[0057]

[0058] After multiple decoding stages, the final features are mapped to reconstruct the infrared image. This output both suppresses weather interference and maintains the continuity of target outline, relative thermal intensity, and local thermal structure.

[0059] In step 8), to avoid unreliable weather semantics and degradation level estimates jointly perturbing the expert route during early training, a two-stage optimization is adopted. The total loss for the first stage is:

[0060] in, For pixel reconstruction loss; Cross-entropy for weather categories; To contrast the recovery loss of prompts, it is used to make the results of correct weather prompts closer to the clear image, and to make the results of incorrect prompts maintain a greater perceptual distance from the clear image.

[0061] The second stage introduces degradation estimation and hierarchical expert routing based on the parameters of the first stage. The total loss is:

[0062] in, Used for aligning degradation prediction with proxy labels; Used to constrain the implicit degree expectation of routing weights and Consistent; Expert collapse can be suppressed by using the coefficient of variation of expert importance and load, and complexity bias can be introduced based on the amount of expert parameters.

[0063] In one specific embodiment, the training images are randomly cropped to 160×160 pixels and input to the main recovery branch, while the degraded images are scaled to 192×192 pixels and input to the weather semantic encoder; random horizontal flipping and rotation enhancement are employed. Four weather-specific task pools are set up, each containing four complexity level experts. The AdamW optimizer is used with an initial learning rate of 2×10⁻⁶. -4 , , Cosine annealing is used, with a minimum learning rate of 1×10⁻⁶. -7, The batch size is 8. The number of negative feedback samples is 3, the weight of the degradation degree supervision loss is 0.02, the weight of the degradation degree consistency loss is 0.01, and the weight of the expert equilibrium loss is 1.0.

[0064] A multi-weather infrared reconstruction dataset can be constructed by combining rain, fog, and snow samples from AWMM-100K, foggy day samples from M3FD-Synthetic infrared data, and thin and thick cloud samples from CUHK-CR1 and CUHK-CR2. Evaluation metrics include peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and learnable perceptual distance (LPIPS), with higher PSNR and SSIM being better, and lower LPIPS being better. The average results of different methods on the multi-weather integrated reconstruction task are shown in Table 1.

[0065]

[0066] As shown in Table 1, this invention achieved an average PSNR of 29.5739 dB, an average SSIM of 0.9213, and an average LPIPS of 0.0922 on the integrated recovery tasks for four weather conditions. Compared with PromptIR, which has the second-best average PSNR and LPIPS, the PSNR is improved by 2.9589 dB and the LPIPS is reduced by 0.0607. Compared with MaIR, which has the second-best average SSIM, the SSIM is improved by 0.0484, indicating that type-degree hierarchical synergy can effectively reduce cross-weather interference and improve thermal structure preservation. The integrated recovery results for each weather sub-task are shown in Table 2.

[0067]

[0068] As shown in Table 2, this invention performs outstandingly in all three metrics for rain and snow removal tasks; it balances structural similarity and perception quality in the fog removal task; and it achieves lower LPIPS and higher SSIM in cloud occlusion recovery, enabling the reconstruction of more continuous roads, buildings, and target edges. A comparison of quantitative results under a single degradation-specific training condition is shown in Table 3.

[0069]

[0070] Table 3 shows that when the present invention is trained for individual degradation, it remains effective for rain removal, fog removal, snow removal, and cloud removal. In particular, it improves PSNR and SSIM and reduces LPIPS in fog removal, snow removal, and cloud removal tasks. This indicates that the proposed module is not only applicable to unified multi-weather models, but can also be used as a general enhancement component for single-weather infrared recovery networks.

[0071] The key innovation of this invention lies in: 1. Using weather type and regional degradation degree as two complementary and hierarchical conditions: first, weather semantics selects a pool of weather-specific experts, and then regional degradation degree selects experts within the pool based on their complexity level, forming a two-dimensional gating mechanism where "type determines direction and degree determines intensity".

[0072] 2. By using the residual amplitude and horizontal / vertical gradient statistics of clear-degraded paired regions to construct proxy labels for regional degradation, and performing weather-conditional normalization, the estimation of spatial non-uniform degradation can be supervised without manual regional labeling.

[0073] 3. The global weather semantic prior and the local feature prior in the decoding stage are used together for multi-level prompt selection, and adaptive prompts are given in the sparse selection stage at different decoding stages, so as to achieve gradual decoupling between weather interference and infrared thermal structure.

[0074] 4. Set up a shared recovery branch and a weather-specific expert branch in each decoding stage: the shared branch recovers weather-independent thermal structures, and the expert branch handles weather-specific patterns and regional degradation differences, and is synergistically fused with the prompt modulation features.

[0075] 5. By matching the bias to the center of degradation degree, the predicted degradation degree value is explicitly mapped to experts with different recovery capabilities; different experts form a complexity gradient from light to heavy with convolution kernel, network depth and parameter size.

[0076] 6. A phased optimization approach is adopted, which involves prior semantic / cue alignment and subsequent degradation estimation / expert routing. Degradation supervision, routing consistency, and complexity-weighted expert loss are also implemented to improve routing stability and prevent expert collapse.

[0077] 7. Extract weather semantics using task-level high-frequency enhanced views, and use bottleneck frequency embedding to assist in degradation degree estimation and expert selection, so that weather categories, local structural damage and recovery paths are coordinated within a unified model.

[0078] In summary, this invention achieves differentiated adaptive restoration of image regions through a hierarchical control strategy that coordinates weather type and regional degradation degree, effectively improving the problems of over-restoration and under-restoration of infrared images. Relying on a multi-level cue-based collaborative decoding optimization method, this invention effectively removes weather degradation interference and infrared thermal structure features, ensuring the detail authenticity and contrast accuracy of the restored image. Simultaneously, this invention employs low-cost proxy tags combined with an expert routing mechanism, significantly reducing annotation costs, suppressing cross-weather negative migration, and stably achieving high-quality infrared image restoration under multiple weather scenarios.

[0079] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for infrared image restoration guided by weather-region degradation under multiple weather conditions, characterized in that, Includes the following steps: Step 1: Obtain an infrared image to be recovered that contains at least one weather degradation caused by rain, fog, snow, or cloud cover; input the infrared image to be recovered into an encoder to obtain multi-scale coding features and bottleneck features; scale the infrared image to be recovered to obtain a task-level view. Step 2: After concatenating the task-level view with its high-frequency response, input the result into the weather semantic encoder to obtain the weather type logical quantity and the global weather semantic prior. The global weather semantic prior is used to limit the prompt selection and the selection of the weather-specific expert pool. Step 3: At each stage of the decoder, generate local cue priors based on the local recovery features of the current stage, calculate the weather conditional cue distribution by combining the global weather semantic priors, sparsely select stage adaptive cue from the cue library of the corresponding stage and fuse it with the features of the current stage to obtain cue modulation features; Deep layer prompts adjust the global recovery direction, mid-layer prompts optimize thermal contrast and main structure, and shallow layer prompts refine edge textures; Step 4: During the training phase, proxy labels for the degree of regional degradation are constructed based on spatially aligned degraded infrared images and clear infrared images. The proxy labels are determined by the regional residual amplitude and residual gradient statistics, and conditional normalization is performed according to weather type to eliminate statistical offsets between different weather categories. Step 5: Based on the semantic information and high-frequency degradation information of the encoder bottleneck features, combined with the global weather semantic prior prediction of the degradation degree score of the region, the prediction results are supervised by the proxy label, so that the model can learn the spatial non-uniform degradation distribution without the need for manual labeling of the region degradation degree. Step 6: Perform multi-source collaborative expert selection at each decoding stage: extract the general thermal structure across weather by sharing the recovery branch, select the target expert pool from multiple weather-specific expert pools by weather semantic priors, and then select experts from multiple complexity levels of experts in the target expert pool that match the current degree of regional degradation according to the degree of regional degradation. Step 7: Fuse the cue modulation features, shared recovery features, and expert recovery features, and generate the features for the next decoding stage step by step through the feedforward enhancement module, and finally output the recovered infrared image; Step 8: Adopt a phased training strategy: The first phase trains the weather semantics, phase prompts, and basic recovery network; The second stage introduces degradation degree estimation, type-degree hierarchical routing, degree consistency, and expert equilibrium constraints to obtain a stable and regionally adaptive multi-weather infrared image restoration model.

2. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 1, the task-level view is obtained from the infrared image to be recovered through bilinear interpolation.

3. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 2, the high-frequency response is obtained by performing a convolution operation on the task-level view using a Laplacian convolution kernel, and the concatenated expression is: in, For task-level view, The symbol represents the Laplacian convolution kernel, and * represents the convolution operation.

4. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 3, the process of generating weather conditional prompt distribution is as follows: global average pooling of the features in the decoding stage is used to obtain the local prompt prior for the stage. After concatenating the global weather semantic prior, a scaling factor is generated through a multilayer perceptron to calibrate the prompt selection logic quantity. A Top-1 sparse strategy is used to select a single stage adaptive prompt, and the feature is fused by a Transformer module and a 1×1 convolution.

5. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 4, the method for constructing the region degradation proxy label is as follows: the image is divided into independent regions, and the residual magnitude between the degraded region and the clear ground truth region is calculated. Horizontal gradient mean Vertical gradient mean Original degree of degradation ; Statistical analysis of the average degree of degradation of similar weather samples Standard deviation The proxy label is obtained after Sigmoid normalization: ; in For gradient weights, The normalized temperature coefficient, is the numerical stability constant.

6. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 5, the degradation representation is obtained by: obtaining a semantic representation of the bottleneck features through global average pooling, and obtaining a frequency embedding through high-pass convolution, activation, pooling and multilayer perceptron. The two are added together to form a degradation representation; the degradation representation is concatenated with the weather semantic prior, input into the degradation degree prediction head, and output the regional degradation degree prediction value.

7. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 6, the specific process of multi-source collaborative expert selection includes: Weather semantic priors are used to perform sparse selection at the task level to determine a dedicated pool of weather experts. Set a degradation level center for each complexity level expert within each weather expert pool; Based on the matching bias between the degree of regional degradation and the center of degradation of each expert, combined with stage features and frequency embedding, the expert routing logic quantity is obtained, and the complexity routing weight is obtained through temperature scaling and sparse selection. A two-dimensional gating system is formed by task-level weights and degradation-level weights to control the output of expert branches.

8. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 7, characterized in that, Experts at different complexity levels form a gradient of recovery capability from light to heavy by adjusting the kernel size, residual stacking depth, and parameter scale.

9. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, In step 8, the two-stage training loss function is set as follows: The total loss in the first phase is: in For pixel reconstruction loss; Cross-entropy for weather categories; To compare and suggest recovery of losses, , These are the weighting coefficients; The total loss in the second phase is: in, To monitor the degree of degradation and loss, For degradation of consistency loss, To balance the losses for the experts, , , These are the weighting coefficients.

10. The method for weather-region degradation collaboratively guided infrared image restoration under multiple weather conditions according to claim 1, characterized in that, The encoder-decoder main network, weather semantic encoder, degradation degree prediction head, multi-level prompt generation module, and multi-level expert pool together constitute the infrared image restoration network. The network can simultaneously process four types of weather degradation: rain, fog, snow, and cloud cover, and output restored infrared images for target detection, segmentation, and recognition visual tasks.