A monitoring image enhancement system and method in a complex meteorological environment based on a visual large model

Abstract

This invention discloses a monitoring image enhancement system and method based on a large visual model under complex meteorological conditions, belonging to the technical field of image enhancement. The system includes a data preprocessing module, a degradation phenomenon recognition model fine-tuning module, a degradation suppression processing module, and a downstream task adaptation module. The data preprocessing module performs spatiotemporal standardization on the monitoring video, unifying the frame rate and resolution, and extracting segments with fixed frame lengths. The degradation phenomenon recognition model fine-tuning module fine-tunes the basic large visual model based on multi-type meteorological degradation datasets, achieving accurate identification of degradation types, degradation regions, and color temperature parameters. The degradation suppression processing module adopts a layered processing strategy to complete targeted removal of meteorological degradation, enhancement of high-frequency residual details, and color and contrast correction. The downstream task adaptation module connects the enhanced image to a downstream visual task model. This invention can simultaneously suppress multiple degradation interferences such as rain, snow, fog, haze, and low light, effectively restoring image details and colors.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and more particularly to a system and method for enhancing monitoring images under complex meteorological environments based on a large visual model. Background Technology

[0002] With the widespread adoption of smart city construction and security monitoring systems, outdoor surveillance cameras have been extensively deployed in urban roads, communities, parks, and transportation hubs. The massive amounts of images and video data collected have become core data support for advanced visual tasks such as intelligent traffic management, target detection, behavior analysis, and public safety early warning. Outdoor surveillance equipment is exposed to the natural environment for extended periods, making its imaging quality highly susceptible to interference from complex weather conditions such as rain, snow, fog, haze, strong light, and low light, leading to severe image degradation. Raindrops and snowflakes can form dynamic, dense, and irregular light spots and obstruction layers on the imaging surface; aerosol particles such as fog and haze can cause a decrease in overall image contrast, blurred details, and a grayish color due to forward and backward scattering; strong midday light or specular reflection can easily cause local overexposure, resulting in the loss of target texture and edge information; and low-light environments present problems such as insufficient brightness, significant noise, and difficulty in feature extraction.

[0003] Existing surveillance image enhancement technologies have significant limitations: traditional image processing methods such as histogram equalization, homomorphic filtering, and dark channel priors have limited effectiveness only for single degradation types like fog and haze. They exhibit poor generalization ability and insufficient detail recovery when facing dynamic, non-uniform, and multi-type complex meteorological interference. While deep learning methods based on convolutional neural networks offer performance improvements over traditional methods, most are designed for single degradation phenomena such as rain, snow, fog, and haze. In practical deployment, different models need to be manually or automatically switched according to weather type, resulting in poor flexibility, cumbersome deployment, high maintenance costs, and insufficient all-weather operational stability. These shortcomings make it difficult to meet the continuous, stable, and high-quality image enhancement requirements of smart city visual systems under complex weather conditions. Summary of the Invention

[0004] To address the challenges of poor image quality and limited adaptability in complex weather conditions, this paper proposes a system and method for enhancing surveillance images under such conditions based on a large visual model. The aim is to achieve the following: First, a single large model can simultaneously suppress image degradation caused by rain, snow, fog, haze, and strong light, overcoming the drawbacks of traditional solutions that require switching models based on weather conditions, thus reducing deployment and maintenance costs. Second, it restores image details, edges, and textures lost due to weather degradation, and corrects color distortion and contrast reduction caused by scattering or overexposure, providing a high-quality data foundation for advanced visual tasks such as intelligent traffic management and public safety early warning.

[0005] The technical solution adopted in this invention is: In a first aspect of the present invention, a monitoring image enhancement system based on a large visual model under complex meteorological conditions is provided, comprising, in sequence: a data preprocessing module, a degradation phenomenon recognition model fine-tuning module, a degradation suppression processing module, and a downstream task adaptation module; The data preprocessing module is used to perform spatiotemporal standardization processing on the original surveillance video, unify the video frame rate and resolution, and cut video segments according to a fixed frame length to form input data blocks of fixed dimensions. The degradation phenomenon recognition model fine-tuning module is used to fine-tune the basic visual model based on a monitoring image dataset containing multiple types of meteorological degradation, so as to obtain a degradation phenomenon recognition model that can identify meteorological degradation types, degradation areas and scene color temperature parameters. The degradation suppression processing module is used to complete the removal of meteorological degradation interference, detail enhancement and optimization, and color and contrast correction by adopting a hierarchical processing strategy based on the degradation features and parameters output by the degradation phenomenon identification model, and outputs a clear image after enhancement. The downstream task adaptation module is used to input the enhanced, clear image into the downstream visual task model to complete the adaptation connection and result output.

[0006] Preferably, the data preprocessing module uses a uniform sampling algorithm to unify the video frame rate to 25fps, uses bilinear interpolation to unify the resolution to 1080p, extracts video segments in units of 16 frames, and fills in segments with less than 16 frames by repeating the last frame.

[0007] Preferably, the degradation phenomenon identification model fine-tuning module includes a data preparation unit, a supervised pair construction unit, and a model training unit; The data preparation unit is used to integrate public traffic datasets and private monitoring data, and to generate various meteorological degradation data such as fog, haze, rain, snow, and low light through physical model simulation. The supervision pair construction unit is used to label the data with degradation type, degradation region and color temperature parameters, and construct a degradation image-normal image supervision pair; The model training unit is used to fine-tune the basic visual model by inputting the supervised dataset, and output a degradation phenomenon recognition model.

[0008] More preferably, the degradation types include fog, haze, rain, snow, and low light; the basic visual large model is CLIP-ViT-L / 14, and the basic visual large model training unit is fine-tuned using the LAMB optimizer, cosine annealing learning rate strategy, and mixed precision training method.

[0009] Preferably, the degradation suppression processing module includes a weather degradation removal submodule, a detail enhancement submodule, and a color contrast correction submodule; The meteorological degradation removal submodule is used to perform corresponding removal operations for different degradation types such as fog, haze, rain, snow and low light. The detail enhancement submodule is used to extract high-frequency features of edge texture through a high-frequency residual compensation network, and to improve resolution and complete residual compensation through sub-pixel convolution. The color contrast correction submodule is used to adjust the hue and saturation in the HSV color space according to the scene color temperature, and to optimize the overall image contrast using an adaptive gain algorithm.

[0010] More preferably, the weather degradation removal submodule performs the following operations: For fog and haze weather, clear images are derived by using atmospheric scattering models to predict transmittance and atmospheric light. For rainy and snowy weather, a dynamic mask generation network is used to obtain rain line masks and snow particle masks, and residual learning is used to remove interference. For low-light weather, the Retinex algorithm is used to decompose the image and an attention mechanism is used to adaptively adjust the illumination and reflection components to achieve brightness balance.

[0011] Preferably, the downstream task adaptation module will enhance the input of the post-image into the pedestrian detection or vehicle detection model without requiring additional format conversion and feature adjustment.

[0012] In another aspect of the present invention, a method for enhancing monitoring images under complex meteorological conditions based on a large visual model is also provided, which is applied to the aforementioned system for enhancing monitoring images under complex meteorological conditions based on a large visual model, and includes the following steps: S1. Perform spatiotemporal standardization preprocessing on the original surveillance video, unify the video frame rate to 25fps and the resolution to 1080p, and extract video segments in fixed 16-frame units to form fixed-dimensional input data blocks; S2. Construct a monitoring image dataset containing multiple meteorological degradation types such as rain, snow, fog, haze, and low light. Based on this dataset, fine-tune the basic visual model to obtain a degradation phenomenon recognition model with the ability to identify meteorological degradation types, degradation areas, and color temperature parameters. S3. Input the preprocessed monitoring video into the fine-tuned degradation phenomenon recognition model, and output the meteorological degradation type, degradation characteristics, transmittance, atmospheric light and scene color temperature parameters. Based on the recognition results, a layered degradation suppression process is adopted, including first removing the corresponding meteorological degradation interference in a targeted manner, then enhancing the details through a high-frequency residual compensation network, and finally completing color and contrast correction in the HSV color space to obtain an enhanced clear image. S4. Input the enhanced image directly into the downstream visual task model to complete the pedestrian detection and vehicle detection adaptation output.

[0013] Compared with the prior art, the present invention has the following significant advantages: 1. Strong scene adaptability and stable operation in all weather conditions: This invention fine-tunes the large visual model through a special dataset. A single model can simultaneously identify and remove multiple meteorological degradations such as rain, snow, fog, haze, and low light. This breaks through the limitations of traditional methods that can only adapt to a single degradation and require switching models according to the weather. It greatly improves the system's generalization ability and all-weather stability in complex meteorological environments and reduces deployment and maintenance costs.

[0014] 2. Superior Enhancement, Complete Restoration of Details and Colors: Employing a layered processing strategy of degradation identification, targeted removal, high-frequency residual enhancement, and color contrast correction, combined with atmospheric scattering models, dynamic masks, Retinex decomposition, and high-frequency residual compensation networks, this approach efficiently removes meteorological interference while accurately restoring details such as image edges, textures, and contours. It also corrects color distortion and contrast reduction caused by scattering and overexposure, resulting in high image fidelity and natural visual effects.

[0015] 3. End-to-end standardization and convenient engineering deployment: Input videos are uniformly standardized in time and space to ensure consistent model input format. The processing flow is automated end-to-end without manual intervention. Enhanced images can be directly input into downstream pedestrian detection and vehicle detection models without format conversion or secondary adjustment. This enables seamless integration between the enhancement module and advanced vision tasks, balancing enhancement accuracy and engineering practicality to meet the needs of large-scale smart city surveillance deployment.

[0016] 4. High target recognition and improved downstream task performance: Through high-frequency feature compensation and adaptive contrast optimization, the feature representation of monitored targets such as pedestrians and vehicles is significantly enhanced, effectively improving the accuracy and robustness of subsequent detection, recognition and tracking tasks, and providing high-quality data support for intelligent transportation and public safety. Attached Figure Description

[0017] Figure 1 This is a framework diagram of a monitoring image enhancement system based on a large visual model under complex meteorological conditions, according to a specific embodiment of the present invention. Figure 2 This is a diagram illustrating the overall technical solution architecture in a specific embodiment of the present invention; Figure 3 This is a flowchart of the degradation phenomenon recognition model fine-tuning module in a specific embodiment of the present invention; Figure 4 This is a flowchart of the degradation suppression processing module in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the high-frequency residual compensation network structure in a specific embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] refer to Figure 1-2 As shown, this embodiment provides a monitoring image enhancement system based on a large visual model under complex meteorological conditions, including: a data preprocessing module, a degradation phenomenon recognition model fine-tuning module, a degradation suppression processing module, and a downstream task adaptation module connected in sequence. The data preprocessing module is used to perform spatiotemporal standardization processing on the original surveillance video, unify the video frame rate and resolution, and cut video segments according to a fixed frame length to form input data blocks of fixed dimensions. In this embodiment, three sets of urban road surveillance videos from different sources were selected: the first set was surveillance video of urban main roads during the morning rush hour (7:30-8:30), with an original frame rate fluctuating between 15-30fps and an original resolution of 720p; the second set was surveillance video of urban secondary roads during the evening rush hour (17:30-18:30), with an original frame rate fluctuating between 20-28fps and an original resolution of 1080p; and the third set was surveillance video of urban side roads at night (22:00-23:00), with an original frame rate fluctuating between 18-25fps and an original resolution of 480p.

[0020] The three video sets were standardized using uniform sampling. For the first set of videos with fluctuating frame rates, one frame was extracted every 1-2 frames to uniformly adjust the frame rate to 25fps. For the second set of videos, missing frames were added according to the original frame rate distribution to maintain 25fps output. The third set of videos was boosted to 25fps using interpolation. The resolution was then uniformly adjusted and scaled to 1080p using bilinear interpolation.

[0021] To ensure the uniformity of the model input format, the processed long video is segmented. Using a fixed unit of 16 frames, continuous segments are extracted from the three groups of videos in sequence to form input data blocks with dimensions [16, 3, 1080, 1920]. The last segment with less than 16 frames is padded by repeating the last frame.

[0022] The degradation phenomenon recognition model fine-tuning module is used to fine-tune the basic visual model based on a monitoring image dataset containing multiple types of meteorological degradation, so as to obtain a degradation phenomenon recognition model that can identify meteorological degradation types, degradation areas and scene color temperature parameters. The degradation phenomenon identification model fine-tuning module includes a data preparation unit, a supervised pair construction unit, and a model training unit; The data preparation unit is used to integrate public traffic datasets and private monitoring data, and to generate various meteorological degradation data such as fog, haze, rain, snow, and low light through physical model simulation. The supervision pair construction unit is used to label the data with degradation type, degradation region and color temperature parameters, and construct a degradation image-normal image supervision pair; The model training unit is used to fine-tune the basic visual model by inputting the supervised dataset, and output a degradation phenomenon recognition model.

[0023] Please see Figure 3 In this embodiment, the data preparation unit constructs a meteorological degradation dataset for fine-tuning the large model. The dataset has a total sample size of 17,000 images, covering monitoring scenarios at different times of day (morning, noon, and evening) and on different road sections of main roads, secondary roads, and secondary roads. It includes five types of meteorological degradation phenomena: fog, haze, rain, snow, and low light. The dataset consists of: 2,000 images related to urban road monitoring selected from the public Cityscapes traffic dataset; 1,500 privately collected urban road monitoring images each from 7:00-9:00, 11:00-13:00, and 17:00-19:00, and 3,000 images from 22:00 to 5:00 the next day; and, based on the images without degradation phenomena in the above data, 1,500 images of each type of degradation (fog, haze, rain, snow, and low light) are simulated using a physical model.

[0024] The supervised pair construction units were labeled using the LabelMe annotation tool. The annotation content included three parts: first, the meteorological degradation type, clearly labeling the fog, haze, rain, snow, and low light category for each image; second, the degradation region, using polygonal boxes to select the fog / haze coverage area, rain line distribution area, or snow particle accumulation area; and third, the scene color temperature parameter, extracting and recording the color temperature value of each image using image analysis tools. After the annotation was completed, each degraded image was paired with its corresponding original clear image to construct a "degraded image-normal image" supervised pair, where the normal image is the original base data without added degradation effects, ultimately forming a complete supervised pair dataset.

[0025] In the model training unit, CLIP-ViT-L / 14 was selected as the basic large-scale vision model. The labeled supervised pair dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model parameter optimization; the validation set was used for dynamically adjusting the learning rate and early stopping strategy; and the test set was used to evaluate the final model performance. Mixed precision training (FP16) was used, with a batch size of 16 and 15 training epochs. The optimizer used was LAMB, with an initial learning rate of 5e-6 and a weight decay coefficient of 0.02. The learning rate decayed every 5 epochs using a cosine annealing strategy. After training, the degradation type and feature extraction effect were validated on the test set, and the weights were saved to output the model.

[0026] The degradation suppression processing module is used to complete the removal of meteorological degradation interference, detail enhancement and optimization, and color and contrast correction by adopting a hierarchical processing strategy based on the degradation features and parameters output by the degradation phenomenon identification model, and outputs a clear image after enhancement. The degradation suppression processing module includes a weather degradation removal submodule, a detail enhancement submodule, and a color contrast correction submodule; The meteorological degradation removal submodule is used to perform corresponding removal operations for different degradation types such as fog, haze, rain, snow and low light. The detail enhancement submodule is used to extract high-frequency features of edge texture through a high-frequency residual compensation network, and to improve resolution and complete residual compensation through sub-pixel convolution. The color contrast correction submodule is used to adjust the hue and saturation in the HSV color space according to the scene color temperature, and to optimize the overall image contrast using an adaptive gain algorithm.

[0027] Please see Figure 4 The meteorological degradation removal submodule performs the following operations: For fog and haze weather, clear images are derived by using atmospheric scattering models to predict transmittance and atmospheric light. For rainy and snowy weather, a dynamic mask generation network is used to obtain rain line masks and snow particle masks, and residual learning is used to remove interference. For low-light weather, the Retinex algorithm is used to decompose the image and an attention mechanism is used to adaptively adjust the illumination and reflection components to achieve brightness balance.

[0028] The meteorological degradation removal submodule targets a degraded image due to fog. By setting the transmittance t=0.4 and atmospheric light A=230 parameters from the fine-tuning model output and substituting them into the atmospheric scattering model, J=(-230×0.6) / 0.4 is obtained, resulting in a clear image J, thus completing the fog removal. For rainy images, the pixel values ​​of the rain line region in the degraded image I are 150-200. Based on the rain line position features output by the model, the predicted value of the rain line mask R corresponding to the region is 1. Through the residual learning mechanism, the pixel values ​​of the rain line region are restored to the corresponding values ​​of the clear image after substituting them into J=IR, thus removing interference. The Retinex image decomposition algorithm is used to decompose the degraded image I into an illumination component L and a reflection component R (I=L×R). By setting the grayscale pixel value <80, the insufficiently illuminated areas are located, and the brightness of the illumination component L in these areas is adjusted to increase the original brightness to 1.5-2.5 times, while keeping the reflection component R unchanged. A balanced image is reconstructed, achieving low-light degradation removal.

[0029] In the detail enhancement submodule, the degraded image J is input into the "high-frequency residual compensation network". Please refer to [link to relevant documentation]. Figure 5 The network consists of four convolutional layers (3×3 kernel size) and one subpixel convolutional layer. The first three convolutional layers extract high-frequency features such as image edges and textures. The subpixel convolutional layer improves the image resolution. The fourth convolutional layer then injects the extracted high-frequency features as residual compensation terms into the high-resolution image to supplement the details lost during the degradation removal process and enhance the feature recognition of the monitored target.

[0030] In the color contrast correction submodule, the enhanced image is converted from the RGB color space to the HSV color space. Based on the scene color temperature parameters predicted by the model, the hue (H channel) and saturation (S channel) are adjusted: when the color temperature is <4000K (warm tone), the H channel value is decreased by 5-10 units, and the S channel value is increased by 10-15%; when the color temperature is >5500K (cool tone), the H channel value is increased by 5-10 units, and the S channel remains unchanged. After adjustment, the image grayscale histogram is calculated. Pixels with grayscale values ​​concentrated between 50-150 receive a gain increase of 1.2-1.5 times, while pixels with grayscale values ​​>230 receive a gain decrease of 0.7-0.9 times, resulting in a more uniform grayscale distribution across the entire image.

[0031] The downstream task adaptation module is used to input the enhanced clear image into the downstream visual task model to complete the adaptation connection and result output. In this embodiment, the video images processed by degradation suppression are directly input into the pedestrian and vehicle detection models without additional format conversion or feature adjustment. Testing showed that the processed video images achieved a pedestrian detection accuracy of 95.2% and a vehicle detection accuracy of 97.1%, demonstrating improved detection performance compared to the unprocessed degraded video (90.3% pedestrian detection accuracy and 88.5% vehicle detection accuracy).

[0032] In a second aspect of the invention, a method for enhancing monitoring images under complex meteorological conditions based on a large visual model is proposed, applied to the aforementioned system for enhancing monitoring images under complex meteorological conditions based on a large visual model, comprising: S1. Perform spatiotemporal standardization preprocessing on the original surveillance video, unify the video frame rate to 25fps and the resolution to 1080p, and extract video segments in fixed 16-frame units to form fixed-dimensional input data blocks; S2. Construct a monitoring image dataset containing multiple meteorological degradation types such as rain, snow, fog, haze, and low light. Based on this dataset, fine-tune the basic visual model to obtain a degradation phenomenon recognition model with the ability to identify meteorological degradation types, degradation areas, and color temperature parameters. S3. Input the preprocessed monitoring video into the fine-tuned degradation phenomenon recognition model, and output the meteorological degradation type, degradation characteristics, transmittance, atmospheric light and scene color temperature parameters. Based on the recognition results, a layered degradation suppression process is adopted, including first removing the corresponding meteorological degradation interference in a targeted manner, then enhancing the details through a high-frequency residual compensation network, and finally completing color and contrast correction in the HSV color space to obtain an enhanced clear image. S4. Input the enhanced image directly into the downstream visual task model to complete the pedestrian detection and vehicle detection adaptation output.

[0033] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A monitoring image enhancement system based on a large visual model under complex meteorological conditions, characterized in that, It includes, in sequence: a data preprocessing module, a degradation phenomenon identification model fine-tuning module, a degradation suppression processing module, and a downstream task adaptation module; The data preprocessing module is used to perform spatiotemporal standardization processing on the original surveillance video, unify the video frame rate and resolution, and cut video segments according to a fixed frame length to form input data blocks of fixed dimensions. The degradation phenomenon recognition model fine-tuning module is used to fine-tune the basic visual model based on a monitoring image dataset containing multiple types of meteorological degradation, so as to obtain a degradation phenomenon recognition model that can identify meteorological degradation types, degradation areas and scene color temperature parameters. The degradation suppression processing module is used to complete the removal of meteorological degradation interference, detail enhancement and optimization, and color and contrast correction by adopting a hierarchical processing strategy based on the degradation features and parameters output by the degradation phenomenon identification model, and outputs a clear image after enhancement. The downstream task adaptation module is used to input the enhanced, clear image into the downstream visual task model to complete the adaptation connection and result output.

2. The image enhancement system for monitoring under complex meteorological environments based on a large visual model according to claim 1, characterized in that, The data preprocessing module uses a uniform sampling algorithm to unify the video frame rate to 25fps, and uses bilinear interpolation to adjust the resolution to 1080p. It extracts video segments in units of 16 frames, and fills in segments with less than 16 frames by repeating the last frame.

3. The image enhancement system for monitoring under complex meteorological environments based on a large visual model according to claim 1, characterized in that, The degradation phenomenon identification model fine-tuning module includes a data preparation unit, a supervised pair construction unit, and a model training unit; The data preparation unit is used to integrate public traffic datasets and private monitoring data, and to generate various meteorological degradation data such as fog, haze, rain, snow, and low light through physical model simulation. The supervision pair construction unit is used to label the data with degradation type, degradation region and color temperature parameters, and construct a degradation image-normal image supervision pair; The model training unit is used to fine-tune the basic visual model by inputting the supervised dataset, and output a degradation phenomenon recognition model.

4. The image enhancement system for monitoring under complex meteorological environments based on a large visual model according to claim 3, characterized in that, The degradation types include fog, haze, rain, snow, and low light; the basic visual model is CLIP-ViT-L / 14, and the basic visual model training unit is fine-tuned using the LAMB optimizer, cosine annealing learning rate strategy, and mixed precision training method.

5. The image enhancement system for monitoring under complex meteorological environments based on a large visual model according to claim 1, characterized in that, The degradation suppression processing module includes a weather degradation removal submodule, a detail enhancement submodule, and a color contrast correction submodule; The meteorological degradation removal submodule is used to perform corresponding removal operations for different degradation types such as fog, haze, rain, snow and low light. The detail enhancement submodule is used to extract high-frequency features of edge texture through a high-frequency residual compensation network, and to improve resolution and complete residual compensation through sub-pixel convolution. The color contrast correction submodule is used to adjust the hue and saturation in the HSV color space according to the scene color temperature, and to optimize the overall image contrast using an adaptive gain algorithm.

6. The image enhancement system for monitoring in complex meteorological environments based on a large visual model according to claim 5, characterized in that, The weather degradation removal submodule performs the following operations: For fog and haze weather, clear images are derived by using atmospheric scattering models to predict transmittance and atmospheric light. For rainy and snowy weather, a dynamic mask generation network is used to obtain rain line masks and snow particle masks, and residual learning is used to remove interference. For low-light weather, the Retinex algorithm is used to decompose the image and an attention mechanism is used to adaptively adjust the illumination and reflection components to achieve brightness balance.

7. The image enhancement system for monitoring under complex meteorological environments based on a large visual model according to claim 1, characterized in that, The downstream task adaptation module will input the enhanced image into the pedestrian detection or vehicle detection model without requiring additional format conversion or feature adjustment.

8. A method for enhancing monitoring images under complex meteorological conditions based on a large visual model, applicable to a monitoring image enhancement system under complex meteorological conditions based on a large visual model as described in any one of claims 1-7, characterized in that, Includes the following steps: S1. Perform spatiotemporal standardization preprocessing on the original surveillance video, unify the video frame rate to 25fps and the resolution to 1080p, and extract video segments in fixed 16-frame units to form fixed-dimensional input data blocks; S2. Construct a monitoring image dataset containing multiple meteorological degradation types such as rain, snow, fog, haze, and low light. Based on this dataset, fine-tune the basic visual model to obtain a degradation phenomenon recognition model with the ability to identify meteorological degradation types, degradation areas, and color temperature parameters. S3. Input the preprocessed monitoring video into the fine-tuned degradation phenomenon recognition model, and output the meteorological degradation type, degradation characteristics, transmittance, atmospheric light and scene color temperature parameters. Based on the recognition results, a layered degradation suppression process is adopted, including first removing the corresponding meteorological degradation interference in a targeted manner, then enhancing the details through a high-frequency residual compensation network, and finally completing color and contrast correction in the HSV color space to obtain an enhanced clear image. S4. Input the enhanced image directly into the downstream visual task model to complete the pedestrian detection and vehicle detection adaptation output.

Images