A weather-aware based monitoring image quality diagnosis method and system

Abstract

The application relates to the technical field of image processing, and particularly discloses a monitoring image quality diagnosis method and system based on weather sensing, which comprises the following steps: collecting monitoring image data from a target monitoring camera; calling a plurality of third-party weather information APIs to obtain real-time weather data of a monitoring area, comparing weather conditions and visibility information provided by the APIs, and confirming the credibility of the weather information when the error is within a preset range; searching for an image classification model that is completely matched in weather condition and closest in visibility in an image quality identification model library according to the verified weather information; identifying quality defects of the monitoring image data by using the matched image classification model, wherein the quality defects include one or more of lens blur, stain pollution and foreign matter shielding; when a quality anomaly is identified, the monitoring image data of the target monitoring camera is acquired again for secondary verification, and if the monitoring image data is still abnormal, fault alarm information is output.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for diagnosing the quality of surveillance images based on weather perception. Background Technology

[0002] With the rapid development of new urban management concepts such as digital twin cities, smart cities, smart transportation, and intelligent security, video surveillance systems have become crucial data collection and sensing terminals in urban infrastructure. By capturing dynamic images of the environment in real time, surveillance cameras not only support public safety and traffic management but also provide multi-dimensional real-world data support for the construction of digital twin cities, including real-time dynamic updates of building status, weather changes, traffic flow, and pedestrian distribution. This data is integrated into the digital twin platform, enabling precise simulation, analysis, and optimization of urban operations, and is an indispensable technological foundation for achieving intelligent urban management and decision-making.

[0003] As a key front-end device in a surveillance system, the quality of camera imaging directly affects the overall effectiveness of the surveillance network and the accuracy of subsequent intelligent analysis (such as target detection and behavior recognition). Therefore, real-time and automated monitoring and diagnosis of surveillance camera image quality, and timely detection of abnormal conditions such as lens blurring, contamination, and obstruction, are of great significance for ensuring the stable operation and data reliability of the surveillance system and digital twin platform.

[0004] Traditional surveillance camera image quality inspection mainly relies on manual inspection and review, which is time-consuming, labor-intensive, and difficult to meet the real-time, continuous monitoring needs of large-scale video surveillance systems. In recent years, the rise of deep learning, especially convolutional neural network (CNN) technology, has made quality assessment based on automated image classification models possible. Most existing methods train a single general-purpose model and use datasets containing normal and various abnormal images to classify and judge surveillance images, achieving automated quality inspection.

[0005] However, existing technologies have significant shortcomings, especially under complex and variable outdoor weather conditions, where the performance of a single general image quality assessment model is often less than ideal. Outdoor weather conditions such as sunny days, cloudy days, rain, snow, fog, haze, and sandstorms can have complex and significant impacts on the visual characteristics of images, causing phenomena such as rain streaks, natural blur, low contrast, graying, or color cast. For example, raindrop interference and overall blurring caused by heavy rain, loss of distant details due to heavy fog, and yellow color cast and granular noise caused by sandstorms—these changes in image features caused by natural environmental factors are highly similar to the abnormal performance caused by camera hardware failures (such as lens defocusing, dirt, or obstruction).

[0006] Because most existing models fail to use real-time weather information as a basis for judgment and lack the ability to perceive environmental changes, they cannot effectively distinguish between image appearance changes caused by weather and actual faults. For example, natural blurring in smoggy weather is often misjudged as lens defocusing; rain blurring is confused with lens smudges; and image sharpness reduction caused by sandstorms is incorrectly identified as equipment failure. Such misjudgments lead to frequent false alarms and false negatives, which not only seriously reduces the accuracy and reliability of automated quality diagnosis systems but also increases the troubleshooting burden on maintenance personnel and affects the timeliness of fault handling.

[0007] The root cause of the aforementioned shortcomings in existing technologies lies in the fact that single, general image quality detection models do not fully consider the dynamic impact of external environmental conditions such as weather on image features, nor do they incorporate real-time weather information to adjust detection strategies and parameters for intelligent adaptation. In other words, these methods adopt a "one-size-fits-all" approach, attempting to cope with diverse and constantly changing environments through a fixed model. They are unable to adjust weights and discrimination criteria for different weather scenarios, resulting in significant accuracy reduction and insufficient robustness in practical applications.

[0008] In summary, given the high requirements of digital twin cities for the accuracy of multi-source real-time data and the trend of intelligent development in monitoring system management, there is an urgent need for a technical solution that can intelligently sense changes in the monitoring environment and weather, and dynamically select the most appropriate image classification model based on real-time weather information. Summary of the Invention

[0009] The purpose of this invention is to design a weather-sensing-based monitoring image quality diagnosis method and system, which solves the problem of insufficient diagnostic accuracy of traditional general methods under different weather conditions. By sensing and utilizing real-time, verified weather information, a dedicated image quality recognition model suitable for the current weather conditions is dynamically matched, thereby achieving more accurate and robust image quality defect identification and diagnosis, reducing false alarms and missed alarms, and improving the reliability and operation and maintenance efficiency of the monitoring system.

[0010] This invention provides a method for diagnosing the quality of surveillance images based on weather perception, comprising:

[0011] S1: Acquire surveillance image data from the target surveillance camera;

[0012] S2: Call several third-party weather information APIs to obtain real-time weather data of the monitored area, compare the weather conditions and visibility information provided by each API, and confirm the credibility of the weather information when the error is within the preset range.

[0013] S3: Based on the verified weather information, retrieve the image classification model in the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility. The image quality recognition model library uses metadata management to store dedicated models for different weather conditions.

[0014] S4: Use the matched image classification model to identify quality defects in the surveillance image data. The quality defects include one or more of lens blur, stains, and foreign object obstruction.

[0015] S5: When a quality anomaly is detected, the monitoring image data of the target surveillance camera is acquired again for secondary verification. If the anomaly is still detected, a fault alarm message is output.

[0016] Preferably, in step S1, acquiring surveillance image data from the target surveillance camera includes:

[0017] The system periodically captures surveillance image data from the target surveillance camera;

[0018] Responding to internal system trigger commands, it proactively captures surveillance image data from the target surveillance camera in real time.

[0019] Preferably, in step S2, the comparison of weather conditions and visibility information provided by each API includes:

[0020] Calculate the standard deviation of the visibility values ​​provided by each API, and confirm the reliability of the visibility information when the standard deviation is less than a preset threshold;

[0021] The weather conditions provided by each API are voted on and statistically analyzed. When the consistency reaches a preset ratio, the weather conditions are confirmed to be reliable.

[0022] Preferably, in step S3, the step of retrieving an image classification model from the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility based on the verified weather information includes:

[0023] The verified weather conditions are used as the exact matching criteria to filter the metadata records; and among the filtered results, the model whose applicable visibility range includes the current real-time visibility, or is closest to the current real-time visibility value is selected.

[0024] The model training process is triggered when no image classification model matching the current weather information is found in the model library.

[0025] Preferably, in step S3, the image quality recognition model library adopts a metadata management method including:

[0026] For each image classification model in the image quality recognition model library, a metadata record is established. The metadata record shall include at least the model's unique identifier, the weather type corresponding to the model, the applicable visibility range, and the storage path of the model file.

[0027] Preferably, the model training process includes:

[0028] The monitoring image data under the current weather conditions is cached as training samples. Quality defect category labels are added to the samples by manual annotation. A new image classification model is trained using the labeled samples. Finally, the trained image classification model and its corresponding weather metadata are stored in the model library.

[0029] Preferably, the image classification model is built based on a pre-trained ResNet-18 network, including a hierarchical feature extraction layer and a feature fusion module;

[0030] The hierarchical feature extraction layer includes the first three feature extraction layers of ResNet-18: layer1, layer2, and layer3, which output feature maps f1, f2, and f3 at different spatial resolutions, respectively.

[0031] The feature fusion module concatenates feature maps f1, f2, and f3 along the channel dimension, performs dimensionality reduction and channel fusion through 1×1 convolution, and then processes them sequentially through batch normalization and ReLU activation function.

[0032] Preferably, in step S4, the step of using the matched image classification model to identify quality defects in the monitoring image data includes:

[0033] The feature fusion module of the image classification model concatenates feature maps from different levels along the channel dimension, and then processes them with the ReLU activation function and global average pooling to output a vector of fixed dimensions. This vector is then fed into Dropout and fully connected layers to classify quality defect categories.

[0034] Preferably, in step S5, the method for outputting fault alarm information includes:

[0035] Alarm notifications containing camera identification, anomaly type, and current weather information are sent to pre-defined maintenance personnel or management systems via the robot interface or message queue of instant messaging tools.

[0036] Preferably, a weather-sensing-based surveillance image quality diagnostic system includes:

[0037] The image information acquisition module is used to collect surveillance image data from the target surveillance camera;

[0038] The weather information acquisition module calls several third-party weather information APIs to obtain real-time weather data of the monitored area, compares the weather conditions and visibility information provided by each API, and confirms the credibility of the weather information when the error is within a preset range.

[0039] The model selection module is used to retrieve an image classification model from the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility, based on the verified weather information.

[0040] The defect identification module is used to identify quality defects in the monitoring image data using the matched image classification model. The quality defects include one or more of lens blur, stains, and foreign object obstruction.

[0041] The secondary verification module is used to re-acquire the monitoring image data of the target surveillance camera for secondary verification when an abnormality is identified. If it is still abnormal, it will output a fault alarm message.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention discloses a weather-sensing-based method and system for diagnosing the quality of surveillance images. By combining real-time weather conditions and selecting corresponding environmental models, it reduces misjudgments caused by weather differences and improves the accuracy and stability of image quality diagnosis. Multi-source weather data verification and secondary image acquisition verification reduce the probability of false alarms and missed alarms, ensuring the reliability of system alarms. The model library, classified and stored based on weather and visibility, allows the system to adapt to changing natural environments and improves the intelligence level of surveillance quality management. The secondary verification and alarm mechanism after anomaly identification ensures timely maintenance, enabling maintenance personnel to quickly locate and handle problems, reducing blind spots and security risks. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a monitoring image quality diagnosis method based on weather perception provided in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of a monitoring image quality diagnosis system module based on weather perception provided in an embodiment of the present invention. Detailed Implementation

[0046] 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.

[0047] like Figure 1As shown, this application provides a weather-sensing-based method for diagnosing the quality of surveillance images, including:

[0048] S1: Acquire surveillance image data from the target surveillance camera;

[0049] S2: Call several third-party weather information APIs to obtain real-time weather data of the monitored area, compare the weather conditions and visibility information provided by each API, and confirm the credibility of the weather information when the error is within the preset range.

[0050] S3: Based on the verified weather information, retrieve the image classification model in the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility. The image quality recognition model library uses metadata management to store dedicated models for different weather conditions.

[0051] S4: Use the matched image classification model to identify quality defects in the surveillance image data. The quality defects include one or more of lens blur, stains, and foreign object obstruction.

[0052] S5: When a quality anomaly is detected, the monitoring image data of the target surveillance camera is acquired again for secondary verification. If the anomaly is still detected, a fault alarm message is output.

[0053] The above scheme integrates real-time weather information with a dynamic model selection mechanism to achieve intelligent diagnosis of surveillance camera image quality. Specifically, it includes five main steps: First, image data is collected from designated surveillance cameras as the basis for diagnosis; second, the reliability of the acquired weather data (such as weather conditions and visibility) is ensured by calling multiple third-party weather APIs and performing cross-validation; third, based on the verified weather information, the image classification model that best matches the current weather conditions is retrieved from a pre-built image quality recognition model library; fourth, the quality defect identification of the collected image data is performed using this matching model, classifying common camera fault types (such as lens blur, stains, or foreign object obstruction); fifth, when the diagnostic results show abnormalities, a secondary image acquisition and verification mechanism is implemented to confirm the authenticity of the fault and output alarm information for timely handling.

[0054] In step S1, real-time image data is acquired from a specific surveillance camera via a network communication protocol (such as HTTP). The acquisition method can be periodic or event-triggered, ensuring that the acquired image data represents the current visual state of the monitored area. The image data, serving as the basis for subsequent diagnostics, is typically stored in a digital format (such as JPEG or RAW) and can be managed through an object storage system.

[0055] In step S2, weather data for the current monitored area is obtained by calling multiple external weather service interfaces, such as Hefeng Weather and OpenWeatherMap. This includes weather conditions (e.g., sunny, cloudy, rainy, snowy) and visibility (in meters or kilometers, quantifying air transparency). Subsequently, data comparison is performed: the consistency of weather condition categories (e.g., string matching) and the numerical differences in visibility (e.g., absolute or relative error) between the APIs are calculated. Data reliability is confirmed only if the differences are within a preset threshold (completely consistent weather conditions, visibility error less than 10%); otherwise, an alarm mechanism is triggered, such as notifying maintenance personnel to replace the API.

[0056] In step S3, the weather data validated in S2 is used as query conditions to search the pre-built model library. The model library uses a relational database to store model metadata (including model ID, weather type, visibility, model path, etc.), while the model files themselves are saved in object storage. The search logic prioritizes ensuring a perfect match of weather conditions (e.g., exact string comparison), then optimizes visibility matching (e.g., minimizing the absolute difference) to select the most suitable model. The image classification model is designed based on a deep learning framework and fine-tuned through transfer learning for specific weather conditions.

[0057] In step S4, the image classification model retrieved in S3 (a custom network based on ResNet) is used to process the image data acquired in S1. The model extracts features through a convolutional neural network and outputs classification results, identifying the categories of image quality defects (such as "no abnormalities," "lens blur," "stains," and "foreign object occlusion"). The output includes predicted labels and confidence scores, with defect types defined based on standard datasets.

[0058] In step S5, if the diagnostic result of S4 shows an anomaly, such as a confidence level higher than the threshold, a secondary verification is triggered: S1 is re-executed to acquire image data, and the model from S4 is applied again for diagnosis. If the secondary result is still abnormal, an alarm message is generated, including fault details, camera ID, and timestamp. The verification mechanism can be set with a delay or multiple iterations to filter out transient interference.

[0059] Preferably, in step S1, acquiring surveillance image data from the target surveillance camera includes:

[0060] The system periodically captures surveillance image data from the target surveillance camera;

[0061] Responding to internal system trigger commands, it proactively captures surveillance image data from the target surveillance camera in real time.

[0062] The above solution supports specifying periods for different surveillance cameras. Capture surveillance images and support the active triggering of surveillance image acquisition from designated cameras by other modules in the system.

[0063] In terms of specific implementation, an IP camera that provides a snapshot interface via the HTTP communication protocol can be used. On the backend, the timed snapshot and active snapshot functions can be further encapsulated using Python / Flask, and corresponding RESTful style interfaces can be provided.

[0064] Preferably, in step S2, the comparison of weather conditions and visibility information provided by each API includes:

[0065] Calculate the standard deviation of the visibility values ​​provided by each API, and confirm the reliability of the visibility information when the standard deviation is less than a preset threshold;

[0066] The weather conditions provided by each API are voted on and statistically analyzed. When the consistency reaches a preset ratio, the weather conditions are confirmed to be reliable.

[0067] In the above solution, multiple third-party weather information APIs can be called to compare whether the information is consistent. The comparison focuses on weather conditions (such as sunny, cloudy, partly cloudy, light rain, heavy rain, thunderstorms, snow, etc.) and visibility (a measure of air transparency, commonly seen in smoggy weather). The weather information is trusted only when the error is within a certain range; otherwise, the operations and maintenance personnel will be notified to replace the API.

[0068] Preferably, in step S3, the step of retrieving an image classification model from the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility based on the verified weather information includes:

[0069] The verified weather conditions are used as the exact matching criteria to filter the metadata records; and among the filtered results, the model whose applicable visibility range includes the current real-time visibility, or is closest to the current real-time visibility value is selected.

[0070] The model training process is triggered when no image classification model matching the current weather information is found in the model library.

[0071] Preferably, in step S3, the image quality recognition model library adopts a metadata management method including:

[0072] For each image classification model in the image quality recognition model library, a metadata record is established. The metadata record shall include at least the model's unique identifier, the weather type corresponding to the model, the applicable visibility range, and the storage path of the model file.

[0073] Preferably, the model training process includes:

[0074] The monitoring image data under the current weather conditions is cached as training samples. Quality defect category labels are added to the samples by manual annotation. A new image classification model is trained using the labeled samples. Finally, the trained image classification model and its corresponding weather metadata are stored in the model library.

[0075] In one embodiment provided in this application, the trained model file is stored as an object using MinIO, and the model's metadata is stored using a relational database (such as MySQL / PostgreSQL). Table 1 is an example of a model metadata representation:

[0076]

[0077] Table 1: Examples of Model Metadata Representation

[0078] Based on the acquired weather information, this module retrieves the most suitable image classification model (which can be fine-tuned based on existing pre-trained models, such as MobileNet) from the existing model library (the weather conditions are exactly the same and the visibility is as close as possible). The retrieval can be achieved through MySQL statements.

[0079] If a suitable model cannot be found, the captured surveillance images are saved and manually labeled. These images can be categorized into types such as no anomalies, blurred images, images with stains or contamination, and images with occlusion, forming a dataset with corresponding weather information for supervised training. Finally, the trained new model is saved into the model library. The dataset format uses a PyTorch-compatible ImageFolder structure, and transfer learning is performed based on PyTorch-provided interfaces (such as defining the loss function via torch.nn.CrossEntropyLoss()).

[0080] Preferably, the image classification model is built based on a pre-trained ResNet-18 network, including a hierarchical feature extraction layer and a feature fusion module;

[0081] The hierarchical feature extraction layer includes the first three feature extraction layers of ResNet-18: layer1, layer2, and layer3 (the layer naming is based on PyTorch), which output feature maps f1, f2, and f3 at different spatial resolutions, respectively.

[0082] The feature fusion module concatenates feature maps f1, f2, and f3 along the channel dimension, performs dimensionality reduction and channel fusion through 1×1 convolution, and then processes them sequentially through batch normalization and ReLU activation function.

[0083] In one embodiment provided in this application, the image classification model is a self-designed convolutional neural network. Based on the existing pre-trained model ResNet, a feature fusion module is added to improve the model's complexity and non-linear modeling capability. This model is suitable for fine-tuning performance on small sample data. The image classification model is implemented using PyTorch, as detailed below:

[0084] First, the first three feature extraction layers of ResNet-18, specifically layers 1, 2, and 3 of torchvision.models.resnet18, are extracted to form the shared foundation of the network. Assume the input image is... The network extracts features through three sub-modules. ,in:

[0085]

[0086] Each level of output feature map has different spatial resolution and semantic expressive power, focusing on edges, textures and local structures respectively.

[0087] To fully leverage the complementary information across layers, a simple yet efficient cross-layer feature fusion module was designed. After concatenating the multi-layer outputs along the channel dimension, a 1×1 convolution is used to achieve dimensionality reduction and channel fusion, i.e.:

[0088]

[0089] in[ [] indicates channel splicing. To merge convolutional kernels, BN stands for Batch Normalization. The ReLU activation function is used. This fusion mechanism not only improves feature utilization but also effectively alleviates the problem of deep overfitting.

[0090] The fused feature maps are then compressed into a fixed-length vector using global average pooling (GAP), and further fed into Dropout and a fully connected layer to output the final classification result. Let the final output be:

[0091]

[0092] The entire network structure balances transfer efficiency and training speed by freezing the bottom layer and training the top layer, making it particularly suitable for rapid modeling of small-scale image datasets.

[0093] Preferably, in step S4, the step of using the matched image classification model to identify quality defects in the monitoring image data includes:

[0094] The feature fusion module of the image classification model concatenates feature maps from different levels along the channel dimension, and then processes them with the ReLU activation function and global average pooling to output a vector of fixed dimensions. This vector is then fed into Dropout and fully connected layers to classify quality defect categories.

[0095] In one embodiment provided in this application, after selecting an image classification model, the module will call the model to classify the captured image (no abnormality, blurred lens, stains, occlusion, etc.). If an abnormality occurs, the module will immediately capture the monitoring image of the corresponding camera again for reclassification. If an abnormality occurs again, the module will notify the maintenance personnel in the same way as the weather information acquisition module to handle it in a timely manner.

[0096] Furthermore, a user-friendly graphical web management interface can be built to support manual review of sampled image classification results. For example, the classification results can be stored in a database, and its table structure can be defined as shown in Table 2:

[0097]

[0098] Table 2: Table structure of classification results

[0099] Preferably, in step S5, the method for outputting fault alarm information includes:

[0100] Alarm notifications containing camera identification, anomaly type, and current weather information are sent to pre-defined maintenance personnel or management systems via the robot interface or message queue of instant messaging tools.

[0101] Preferably, a weather-sensing-based surveillance image quality diagnostic system includes:

[0102] The image information acquisition module is used to collect surveillance image data from the target surveillance camera;

[0103] The weather information acquisition module calls several third-party weather information APIs to obtain real-time weather data of the monitored area, compares the weather conditions and visibility information provided by each API, and confirms the credibility of the weather information when the error is within a preset range.

[0104] The model selection module is used to retrieve an image classification model from the image quality recognition model library that perfectly matches the weather conditions and has the closest visibility, based on the verified weather information.

[0105] The defect identification module is used to identify quality defects in the monitoring image data using the matched image classification model. The quality defects include one or more of lens blur, stains, and foreign object obstruction.

[0106] The secondary verification module is used to re-acquire the monitoring image data of the target surveillance camera for secondary verification when an abnormality is identified. If it is still abnormal, it will output a fault alarm message.

[0107] The above-described solution, through its implementation, can produce the following beneficial technical effects:

[0108] Improving the reliability of weather information acquisition: By calling several third-party weather information APIs and comparing and verifying them in the weather information acquisition module, the accuracy of the weather information used as the basis for subsequent model selection is ensured. When the data from different APIs differs too much, the system can detect potential problems with the weather data source, avoiding deviations in the entire diagnostic process caused by a single, incorrect weather information input, and laying a solid foundation for accurate model matching in the future.

[0109] To enhance adaptability, the diagnostic model selection module dynamically optimizes and selects the model that best matches the weather conditions and has the closest visibility, based on reliable weather information provided by the weather information acquisition module. This means the system no longer uses a single, general-purpose model to handle all weather conditions, but instead adopts the specialized model that performs optimally under specific weather conditions (e.g., heavy rain with 500 meters of visibility or sunny skies with 10 kilometers of visibility). This allows the diagnostic method to dynamically adapt to environmental changes, significantly improving its ability to understand and process image features under different weather backgrounds.

[0110] Improving the accuracy of image quality defect identification and reducing false alarm and false negative rates: The defect identification module employs an image classification model highly matched to current weather conditions. This model, during training, has learned the characteristics of normal images under corresponding weather conditions as well as various fault features. Therefore, it can more effectively distinguish which image appearance changes (such as blurring, decreased contrast, and color cast) are normal phenomena caused by current weather conditions (such as fog, rain, and dust), and which are caused by real faults such as lens blur, dirt, and obstruction. This directly leads to an improved accuracy in identifying real faults and significantly reduces the occurrence of misclassifying normal images as faults (false alarms) or ignoring real faults under adverse weather conditions (false negatives).

[0111] Enhancing the confidence of fault alarms: The secondary verification module introduces a secondary verification mechanism that, after initially identifying an anomaly, re-acquires images and re-identifies the fault to confirm its persistence. This helps filter out short-term image anomalies caused by transient interference (such as birds flying by, brief strong light reflections, temporary vehicle obstruction, etc.), ensuring that alarms are only triggered by persistent conditions that are highly likely to be genuine hardware faults or stable obstructions. This further improves the accuracy and reliability of alarms and reduces unnecessary maintenance responses.

[0112] In summary, this application overcomes the shortcomings of existing technologies where a single model is difficult to adapt to complex weather changes by introducing mechanisms such as weather perception, multi-source weather information verification, dynamic model matching, and secondary verification. This enables high-precision and high-reliability intelligent diagnosis of monitoring image quality under various weather conditions, ultimately improving the operation and maintenance efficiency and overall effectiveness of the monitoring system.

[0113] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A weather-aware based monitoring image quality diagnosis method, characterized in that, Comprise: S1: collect monitoring image data from the target monitoring camera; S2: call several third-party weather information APIs to obtain real-time weather data of the monitoring area, compare the weather conditions and visibility information provided by each API, and confirm the credibility of the weather information when the error is within the preset range; In step S2, the comparison of the weather conditions and visibility information provided by each API comprises: Calculate the standard deviation of the visibility values provided by each API, and confirm the visibility information as credible when the standard deviation is less than a preset threshold; Vote on the weather conditions provided by each API, and confirm the weather conditions as credible when the consistency reaches a preset proportion; S3: According to the verified weather information, search for an image classification model in the image quality recognition model library that completely matches the weather condition and has the closest visibility, and the image quality recognition model library stores special models under different weather conditions using metadata management; In step S3, the image quality recognition model library uses metadata management, which comprises: Use the verified weather condition as the exact matching condition to filter in the metadata record, and select the model whose applicable visibility range contains the current real-time visibility or is closest to the current real-time visibility value from the filtered results; When no image classification model matching the current weather information can be searched in the model library, trigger the model training process; The model training process comprises: Cache the monitoring image data under the current weather condition as a training sample, add quality defect category labels to the sample through manual labeling, train a new image classification model using the labeled sample, and finally store the trained image classification model and its corresponding weather metadata in the model library; The image classification model is constructed based on a pre-trained ResNet-18 network, comprising a hierarchical feature extraction layer and a feature fusion module; The hierarchical feature extraction layer comprises the first three feature extraction layers layer1, layer2, and layer3 of ResNet-18, which respectively output feature maps f1, f2, and f3 with different spatial resolutions; The feature fusion module concatenates the feature maps f1, f2, and f3 in the channel dimension, performs dimension reduction and channel fusion through 1x1 convolution, and then processes them through batch normalization and ReLU activation function in turn; S4: Use the matched image classification model to identify quality defects in the monitoring image data, which include one or more of lens blur, stain pollution, and foreign object obstruction; S5: When a quality anomaly is identified, collect the monitoring image data of the target monitoring camera again for secondary verification, and output a fault alarm information if it is still abnormal.

2. The weather-aware monitoring image quality diagnostic method according to claim 1, wherein, In step S1, the monitoring image data is collected from the target monitoring camera, which comprises: Grab the monitoring image data from the target monitoring camera at a preset periodic time; Respond to the trigger instruction inside the system to actively grab the monitoring image data from the target monitoring camera in real time. 3.The weather-aware monitoring image quality diagnostic method of claim 1, wherein, In step S3, the image quality recognition model library uses metadata management, which comprises: A metadata record is established for each image classification model in the image quality recognition model library, and the metadata record at least contains a model unique identifier, a weather type corresponding to the model, an applicable visibility range, and a storage path of a model file. 4.The weather-aware monitoring image quality diagnostic method of claim 1, wherein, In step S4, the matched image classification model is used to identify quality defects of the monitoring image data, including: The feature fusion module of the image classification model splices the feature maps of different levels in the channel dimension, processes them through a ReLU activation function and a global average pooling, and outputs a fixed-dimensional vector, which is further sent to a Dropout and a fully connected layer for classifying quality defect categories. 5.The weather-aware monitoring image quality diagnostic method of claim 1, wherein, In step S5, the manner of outputting the fault alarm information includes: An alarm notification containing the camera identifier, the abnormal type, and the current weather information is sent to the preset operation and maintenance personnel or management system through a robot interface or a message queue of an instant messaging tool.

Images