Water conservancy safety monitoring system based on dynamic vision

Abstract

The invention belongs to the technical field of image contrast, particularly relates to a water conservancy safety monitoring system based on dynamic vision, and aims to solve the problems that in an existing water conservancy safety monitoring process, the image quality and the recognition accuracy are remarkably reduced under complex and dynamically changing illumination and environment conditions, and the image quality is poor. In order to solve the problems of unstable monitoring performance and insufficient reliability caused by the lack of adaptive optimization and multi-source information deep fusion capability, the invention provides the following scheme: the system comprises a data acquisition module, the data acquisition module is connected with an illumination evaluation module, the illumination evaluation module is connected with a scene classification module, and the scene classification module is connected with a data processing module. According to the invention, through a dynamic vision technology, adaptive image enhancement and multi-source data fusion of a water conservancy project under complex illumination and environment conditions are realized, the image quality and defect identification accuracy of all-weather monitoring are significantly improved, and the environmental adaptability and reliability of the system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of image contrast technology, and in particular to a water conservancy safety monitoring system based on dynamic vision. Background Technology

[0002] Safety monitoring of water conservancy projects is a crucial link in ensuring the long-term stable operation of water conservancy facilities. Traditional water conservancy safety monitoring mainly relies on manual inspections and fixed-point sensor deployment, which has problems such as limited monitoring range, insufficient real-time performance, susceptibility to environmental interference, and difficulty in achieving all-weather automated monitoring. With the development of visual sensing technology, image-based monitoring methods have gradually been applied, but existing technologies still have the following shortcomings: Poor environmental adaptability: The imaging quality of traditional visual monitoring systems deteriorates significantly under complex lighting conditions, rain, fog, and nighttime, affecting the accuracy of target recognition and defect detection; Insufficient fusion of multi-source data: Visible light, infrared and other multi-spectral data are often processed independently, lacking effective spatiotemporal alignment and feature fusion mechanisms, making it difficult to leverage the complementary advantages of multimodal data; Limited level of intelligence: Most systems still rely on image processing algorithms with fixed thresholds, lack scene-based adaptive enhancement and recognition capabilities, and cannot adapt to dynamically changing water conservancy monitoring scenarios.

[0003] In existing technologies, during water conservancy safety monitoring, image quality and recognition accuracy significantly decrease when faced with complex and dynamically changing lighting and environmental conditions. Furthermore, the lack of adaptive optimization and deep fusion capabilities of multi-source information leads to unstable monitoring performance and insufficient reliability. To address these issues, we propose a water conservancy safety monitoring system based on dynamic vision. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies in water conservancy safety monitoring, such as a significant decrease in image quality and recognition accuracy when facing complex and dynamically changing lighting and environmental conditions, and the lack of adaptive optimization and deep fusion capabilities of multi-source information, resulting in unstable monitoring performance and insufficient reliability. Therefore, this invention proposes a water conservancy safety monitoring system based on dynamic vision.

[0005] The water conservancy safety monitoring system based on dynamic vision provided in this application adopts the following technical solution: A water conservancy safety monitoring system based on dynamic vision includes a data acquisition module, which is connected to an illumination assessment module, which is connected to a scene classification module, which is connected to a region segmentation module, and which is connected to a contrast enhancement module. A quality assessment module is connected to a contrast enhancement module, a dynamic adjustment module is connected to a fusion optimization module, and a feature alignment and fusion module is connected to a feature alignment and fusion module. A deep learning recognition module is connected to a fusion optimization module and a feature alignment and fusion module. The deep learning recognition module is also connected to a comparison analysis module, which is in turn connected to a reconstruction module. An alarm generation module is connected to a comparison analysis module and a deep learning recognition module. The alarm generation module is also connected to a report generation module, which is connected to a correlation analysis module. The feedback optimization module is connected to the scene classification module and the alarm generation module. The feedback optimization module is also connected to the scheduling optimization module, which in turn is connected to the adaptive update module.

[0006] Furthermore, the data acquisition module includes a multispectral image acquisition unit, a high dynamic range imaging unit, and an environmental parameter synchronous sensing unit. The multispectral image acquisition unit is connected to the high dynamic range imaging unit, and the high dynamic range imaging unit is connected to the environmental parameter synchronous sensing unit. The data acquisition module is responsible for the synchronous and high-quality acquisition of raw monitoring data, providing a multi-dimensional information source for subsequent processing.

[0007] Furthermore, the scene classification module includes an illumination pattern discrimination unit, a weather interference assessment unit, and a scene semantic preliminary judgment unit. The illumination pattern discrimination unit is connected to the weather interference assessment unit, and the weather interference assessment unit is connected to the scene semantic preliminary judgment unit. The scene classification module is used to intelligently identify and classify the current monitoring scene to determine the basic strategy for subsequent image processing.

[0008] Furthermore, the region segmentation module includes a region of focus generation unit, a texture analysis unit, and an edge optimization unit. The region of focus generation unit is connected to the texture analysis unit, and the texture analysis unit is connected to the edge optimization unit. The region segmentation module is used to perform refined and structured region division on the monitoring image.

[0009] Furthermore, the contrast enhancement module includes a spatial enhancement unit, a frequency domain illumination correction unit, and a mapping fusion unit. The spatial enhancement unit is connected to the frequency domain illumination correction unit, and the frequency domain illumination correction unit is connected to the mapping fusion unit. The contrast enhancement module is used to maximize the visual separability of target features and background in response to the characteristics of different regions and the problem of insufficient contrast. The spatial enhancement unit mainly enhances contrast by adjusting the pixel values ​​of the image, using a histogram equalization algorithm; Histogram equalization is a process that adjusts the histogram distribution of an image to make the pixel values ​​more evenly distributed. The steps include: Calculate the histogram of the image ,in Represents the pixel grayscale value. Indicates grayscale value The number of pixels; Calculate the cumulative histogram : ; The cumulative histogram is normalized to the range [0, 255] to obtain new grayscale values. : ; in, The total number of pixels in the image; The grayscale values ​​in the original image are... The pixels are replaced with new grayscale values. This yields an image with enhanced contrast. The frequency domain illumination correction unit mainly processes the image in the frequency domain to correct the effect of uneven illumination on image contrast, using a high-pass filtering algorithm. High-pass filtering algorithms enhance image contrast by filtering out low-frequency components (such as slowly changing parts caused by uneven lighting) and preserving high-frequency components (such as edges and details). The steps include: To transform an image from the spatial domain to the frequency domain, the Fast Fourier Transform (FFT) is typically used. ; in, These are the pixel values ​​of the image. It is its frequency domain representation; Design a high-pass filter. Common high-pass filters include ideal high-pass filters and Butterworth high-pass filters. Taking the ideal high-pass filter as an example, its definition is: ; in, It is a frequency point Distance to the center of the frequency domain It is the cutoff frequency; Multiplying the filter by the frequency domain representation of the image yields the filtered frequency domain image: ; To convert the filtered frequency domain image back to the spatial domain, the inverse fast Fourier transform (IFFT) is typically used. ; in, It is an image with enhanced contrast; The mapping fusion unit fuses the images after spatial domain enhancement and frequency domain illumination correction to obtain better contrast enhancement effect, using a weighted average fusion algorithm; The weighted average fusion algorithm fuses multiple images by adding them together according to certain weights. The steps include: Let the spatially enhanced image be The image after frequency domain illumination correction is ; Choose appropriate weights and ,satisfy ; Calculate the fused image ; ; By adjusting the weights and This allows for the control of the impact of spatial domain enhancement and frequency domain illumination correction on the final image, thereby achieving the best contrast enhancement effect.

[0010] Furthermore, the feature alignment and fusion module includes a spatiotemporal registration unit, a feature-level fusion unit, and a saliency enhancement unit. The spatiotemporal registration unit is connected to the feature-level fusion unit, and the feature-level fusion unit is connected to the saliency enhancement unit. The feature alignment and fusion module is used to achieve the unification and complementarity of multi-source heterogeneous data (visible light and thermal infrared) at the spatiotemporal and feature levels, providing more comprehensive information for identification.

[0011] Furthermore, the deep learning recognition module includes a multi-task recognition unit, an enhanced detection unit, and an inference unit. The multi-task recognition unit is connected to the enhanced detection unit, and the enhanced detection unit is connected to the inference unit. The deep learning recognition module is used to automatically, accurately, and in parallel detect and identify various typical defects and anomalies of water conservancy projects on images that integrate multi-dimensional features, based on a trained deep neural network model.

[0012] Furthermore, the alarm generation module includes a multi-level alarm rule unit, an alarm evaluation unit, and an alarm information generation unit. The multi-level alarm rule unit is connected to the alarm evaluation unit, and the alarm evaluation unit is connected to the alarm information generation unit. The alarm generation module is used to transform the analysis and identification results into standardized and operable early warning information and trigger subsequent processes.

[0013] Furthermore, the feedback optimization module includes a strategy parameter optimization unit, a dynamic scheduling unit, and a model fine-tuning unit. The strategy parameter optimization unit is connected to the dynamic scheduling unit, and the dynamic scheduling unit is connected to the model fine-tuning unit. The feedback optimization module is used to realize the system's self-learning and continuous improvement.

[0014] Furthermore, the correlation analysis module includes a spatiotemporal coupling correlation analysis unit, a multiphysics causal inference unit, and a risk simulation and early warning unit. The spatiotemporal coupling correlation analysis unit is connected to the multiphysics causal inference unit, and the multiphysics causal inference unit is connected to the risk simulation and early warning unit. The correlation analysis module is used to explore the intrinsic relationship between different alarm events and monitoring parameters.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This solution, through its illumination assessment and scene classification modules, can intelligently identify the current environmental conditions and adaptively select and adjust image enhancement strategies, significantly improving image quality and target visibility under adverse conditions such as backlight, haze, and nighttime, thereby increasing recognition accuracy; 2. This solution achieves accurate registration and complementary fusion of multi-source data such as visible light and thermal infrared at the spatiotemporal and feature levels through the feature alignment and fusion module. It not only provides more comprehensive target information (such as combining visible light texture with infrared temperature anomalies), but also enhances the system robustness when a single sensor fails or its performance degrades. 3. This solution uses a contrast enhancement module for adaptive image enhancement, which can effectively overcome the problems of insufficient contrast and blurred details caused by adverse conditions such as backlight, haze, and low light at night, providing high-quality and highly visible input images for core recognition tasks.

[0016] This invention utilizes dynamic vision technology to achieve adaptive image enhancement and multi-source data fusion for water conservancy projects under complex lighting and environmental conditions, significantly improving image quality and defect identification accuracy for all-weather monitoring, and enhancing the system's environmental adaptability and reliability. Attached Figure Description

[0017] Figure 1 This is a structural block diagram of a water conservancy safety monitoring system based on dynamic vision proposed in this invention; Figure 2 This is a structural block diagram of a data acquisition module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention; Figure 3 This is a structural block diagram of a scene classification module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 4This is a structural block diagram of a region segmentation module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 5 This is a structural block diagram of a contrast enhancement module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 6 This is a structural block diagram of a feature alignment and fusion module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 7 This is a structural block diagram of a deep learning recognition module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 8 This is a structural block diagram of an alarm generation module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 9 This is a structural block diagram of a feedback optimization module for a dynamic vision-based water conservancy safety monitoring system proposed in this invention. Figure 10 This is a structural block diagram of the correlation analysis module of a water conservancy safety monitoring system based on dynamic vision proposed in this invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1 Reference Figures 1-10 A water conservancy safety monitoring system based on dynamic vision includes a data acquisition module, a light assessment module connected to the data acquisition module, a scene classification module connected to the light assessment module, a region segmentation module connected to the scene classification module, and a contrast enhancement module connected to the region segmentation module. The quality assessment module is connected to the contrast enhancement module. The quality assessment module is connected to the dynamic adjustment module. The dynamic adjustment module is connected to the fusion optimization module. The fusion optimization module is connected to the feature alignment and fusion module. The deep learning recognition module is connected to the fusion optimization module and the feature alignment and fusion module. The deep learning recognition module is connected to the comparison analysis module, and the comparison analysis module is connected to the reconstruction module. The alarm generation module is connected to the comparison analysis module and the deep learning recognition module. The alarm generation module is connected to the report generation module, and the report generation module is connected to the correlation analysis module. The feedback optimization module is connected to the scene classification module and the alarm generation module. The feedback optimization module is connected to the scheduling optimization module, and the scheduling optimization module is connected to the adaptive update module.

[0020] In this embodiment, the data acquisition module includes a multispectral image acquisition unit, a high dynamic range imaging unit, and an environmental parameter synchronous sensing unit. The multispectral image acquisition unit is connected to the high dynamic range imaging unit, and the high dynamic range imaging unit is connected to the environmental parameter synchronous sensing unit. The data acquisition module is responsible for the synchronous and high-quality acquisition of raw monitoring data, providing a multi-dimensional information source for subsequent processing. The region segmentation module includes a region of focus generation unit, a texture analysis unit, and an edge optimization unit. The region of focus generation unit is connected to the texture analysis unit, and the texture analysis unit is connected to the edge optimization unit. The region segmentation module is used to refine and structure the monitoring images. The system is divided into regions. The deep learning recognition module includes a multi-task recognition unit, an enhanced detection unit, and an inference unit. The multi-task recognition unit is connected to the enhanced detection unit, and the enhanced detection unit is connected to the inference unit. The deep learning recognition module is used to automatically, accurately, and in parallel detect and identify various typical defects and anomalies of water conservancy projects on images that integrate multi-dimensional features, based on a trained deep neural network model. The feedback optimization module includes a strategy parameter optimization unit, a dynamic scheduling unit, and a model fine-tuning unit. The strategy parameter optimization unit is connected to the dynamic scheduling unit, and the dynamic scheduling unit is connected to the model fine-tuning unit. The feedback optimization module is used to realize the system's self-learning and continuous improvement.

[0021] In this embodiment, the contrast enhancement module includes a spatial enhancement unit, a frequency domain illumination correction unit, and a mapping fusion unit. The spatial enhancement unit is connected to the frequency domain illumination correction unit, and the frequency domain illumination correction unit is connected to the mapping fusion unit. The contrast enhancement module is used to maximize the visual separability of target features and background in response to the characteristics of different regions and the problem of insufficient contrast. The spatial enhancement unit mainly enhances contrast by adjusting the pixel values ​​of the image, using a histogram equalization algorithm; Histogram equalization is a process that adjusts the histogram distribution of an image to make the pixel values ​​more evenly distributed. The steps include: Calculate the histogram of the image ,in Represents the pixel grayscale value. Indicates grayscale value The number of pixels; Calculate the cumulative histogram : ; The cumulative histogram is normalized to the range [0, 255] to obtain new grayscale values. : ; in, The total number of pixels in the image; The grayscale values ​​in the original image are... The pixels are replaced with new grayscale values. This yields an image with enhanced contrast. The frequency domain illumination correction unit mainly processes the image in the frequency domain to correct the effect of uneven illumination on image contrast, using a high-pass filtering algorithm. High-pass filtering algorithms enhance image contrast by filtering out low-frequency components (such as slowly changing parts caused by uneven lighting) and preserving high-frequency components (such as edges and details). The steps include: To transform an image from the spatial domain to the frequency domain, the Fast Fourier Transform (FFT) is typically used. ; in, These are the pixel values ​​of the image. It is its frequency domain representation; Design a high-pass filter. Common high-pass filters include ideal high-pass filters and Butterworth high-pass filters. Taking the ideal high-pass filter as an example, its definition is: ; in, It is a frequency point Distance to the center of the frequency domain It is the cutoff frequency; Multiplying the filter by the frequency domain representation of the image yields the filtered frequency domain image: ; To convert the filtered frequency domain image back to the spatial domain, the inverse fast Fourier transform (IFFT) is typically used. ; in, It is an image with enhanced contrast; The mapping fusion unit fuses the images after spatial domain enhancement and frequency domain illumination correction to obtain better contrast enhancement effect, using a weighted average fusion algorithm; The weighted average fusion algorithm fuses multiple images by adding them together according to certain weights. The steps include: Let the spatially enhanced image be The image after frequency domain illumination correction is ; Choose appropriate weights and ,satisfy ; Calculate the fused image ; ; By adjusting the weights and This allows for the control of the impact of spatial domain enhancement and frequency domain illumination correction on the final image, thereby achieving the best contrast enhancement effect.

[0022] In this embodiment, the scene classification module includes an illumination pattern discrimination unit, a weather interference assessment unit, and a scene semantic preliminary judgment unit. The illumination pattern discrimination unit is connected to the weather interference assessment unit, and the weather interference assessment unit is connected to the scene semantic preliminary judgment unit. The scene classification module is used to intelligently identify and classify the current monitoring scene to determine the basic strategy for subsequent image processing. The feature alignment and fusion module includes a spatiotemporal registration unit, a feature-level fusion unit, and a saliency enhancement unit. The spatiotemporal registration unit is connected to the feature-level fusion unit, and the feature-level fusion unit is connected to the saliency enhancement unit. The feature alignment and fusion module is used to achieve the unification and complementarity of multi-source heterogeneous data (visible light and thermal infrared) at the spatiotemporal and feature levels, providing a basis for identification. To provide more comprehensive information, the alarm generation module includes a multi-level alarm rule unit, an alarm evaluation unit, and an alarm information generation unit. The multi-level alarm rule unit is connected to the alarm evaluation unit, and the alarm evaluation unit is connected to the alarm information generation unit. The alarm generation module is used to transform the analysis and identification results into standardized and actionable early warning information and trigger subsequent processes. The correlation analysis module includes a spatiotemporal coupling correlation analysis unit, a multiphysics causal inference unit, and a risk simulation and early warning unit. The spatiotemporal coupling correlation analysis unit is connected to the multiphysics causal inference unit, and the multiphysics causal inference unit is connected to the risk simulation and early warning unit. The correlation analysis module is used to explore the inherent relationships between different alarm events and monitoring parameters.

[0023] The implementation principle in this embodiment is as follows: During use, after the data acquisition module is activated, its internal multispectral image acquisition unit (typically including visible light and thermal infrared cameras) and high dynamic range (HDR) imaging unit work together to capture images from fixed or pan-tilt-zoom (PTZ) monitoring points at the same timestamp. The environmental parameter synchronous sensing unit (such as a light sensor and a thermometer / hygrometer) synchronously records the physical environment data of the site. All data is tagged with a unified time-space label to ensure spatiotemporal consistency for subsequent fusion. The acquired raw visible light / HDR image stream is sent in real-time to the illumination evaluation module and scene classification module. Simultaneously, the thermal infrared image stream and environmental parameter data are cached for subsequent fusion. The illumination evaluation module receives the visible light image. By calculating quantitative indicators such as overall image contrast, brightness distribution (histogram), and shadow area proportion, the current image quality is initially assessed, outputting status signals such as "low contrast," "overexposed," and "underexposed." The scene classification module integrates the illumination assessment results, environmental parameters (such as ambient light intensity, temperature, and humidity), and the semantic features of the image itself (through lightweight neural networks or feature analysis) to accurately classify the current monitoring scene. For example, it may classify it into specific categories such as "daytime dam backlight monitoring," "nighttime low illumination on the water surface," "gate area under haze weather," or "normal illumination on a sunny day." This classification result serves as the "master switch" for all subsequent processing strategy selections. The region segmentation module activates corresponding presets or... Dynamically generated region segmentation strategies, for example in the "dam monitoring" scenario, involve a focus region generation unit that uses prior knowledge (dam outline model) or image segmentation algorithms (such as semantic segmentation based on deep learning) to mark structural regions such as the dam surface, joints, and drainage holes in the image as "focus regions," while marking the sky, water surface, and vegetation as "background regions." Texture analysis and edge optimization units then refine the boundaries of the focus regions to ensure accurate segmentation. The contrast enhancement module receives the segmented image and region mask, and its internal spatial enhancement unit (e.g., using adaptive histogram equalization (CLAHE) instead of global histogram equalization to avoid noise amplification) and frequency domain illumination correction unit (e.g., using...)... The parameters of the Butterworth high-pass filter (using an adaptive cutoff frequency) are dynamically configured based on scene classification (e.g., "backlight" emphasizes shadow enhancement, "haze" emphasizes high-frequency enhancement) and region type (aggressive enhancement is used for key areas, and mild enhancement is used for background areas). The mapping and fusion unit dynamically adjusts the weights α and β in the weighted average algorithm based on feedback from the quality assessment module. For example, at night, frequency domain correction (enhancing details) may be given higher weight; under strong daylight, spatial domain enhancement (adjusting dynamic range) may be more relied upon. The quality assessment module evaluates the image output by the contrast enhancement module using no-reference image quality evaluation metrics (such as BRISQUE, NIQE) or specific metrics for target visibility.If the evaluation score is lower than the preset threshold for that scene category, the dynamic adjustment module is triggered. Based on the quality evaluation results, the dynamic adjustment module generates adjustment instructions, including: fine-tuning the parameters of the contrast enhancement algorithm (such as CLAHE's cropping limit and filter cutoff frequency), suggesting switching to a backup enhancement algorithm combination, and even feeding back to the scene classification module to request a re-evaluation of the scene (when the initial classification may be inaccurate). This process forms a fast internal closed loop, ensuring that the image quality input downstream always meets the requirements. The fusion optimization module first performs preliminary fusion (such as weighted average or PCA fusion) on the data-aligned visible light enhanced image and thermal infrared image to generate a preliminary multispectral image. The spatiotemporal registration unit utilizes the time of acquisition... By using synchronization information and pre-calibrated spatial transformation parameters (or real-time feature point matching), the parallax problem caused by different physical positions of the cameras is solved, achieving pixel-level alignment between visible light and thermal infrared images. The feature-level fusion unit then extracts high-level semantic features from the aligned dual-channel images (e.g., through two parallel convolutional neural network branches). These features are then concatenated along the channel dimension or fused using more complex attention mechanisms. Based on this, the saliency enhancement unit calculates the saliency map of the fused features, highlighting anomalous regions that may contain defects (e.g., high-temperature areas in thermal images corresponding to texture anomalies in visible light images), suppressing irrelevant background, and generating a "comprehensive feature map" rich in multimodal complementary information. Deep learning then identifies the modal... The block receives a "comprehensive feature map" as input. Its multi-task recognition unit loads a pre-trained multi-task deep learning model (such as a multi-head network based on YOLO or Faster R-CNN frameworks). This model can output the results of multiple tasks in parallel during a forward propagation process: including the bounding box location of the defect (localization), the defect category (such as cracks, seepage, erosion, vegetation invasion - classification), and the corresponding confidence score. The enhancement detection unit uses feature pyramids or context information enhancement mechanisms to improve the detection rate for common and difficult-to-detect small target or weak texture defects (such as fine cracks) in water conservancy scenarios. The inference unit performs the forward computation of the model, and the comparison analysis module compares the recognition results of the current frame (defect location, size) with the stored data. The system compares historical monitoring results from the same angle stored in the database, and determines whether the defect is "newly appeared," "expanded," "stable," or "repaired" by calculating the overlap (IOU) and area change rate. The reconstruction module uses difference analysis or image registration technology to visualize the changed area and generate a "defect evolution heatmap" or an image overlaid with change vectors to intuitively show the development trend of the defect. The alarm generation module's multi-level alarm rule unit has a built-in rule base based on business knowledge. The alarm evaluation unit substitutes the identification results and change analysis results into the rule base for evaluation. The alarm information generation unit then generates structured alarm information, including time, location, defect type, level, snapshot, and evolution data, and triggers notifications (audio-visual, SMS, platform pop-up).The report generation module periodically (e.g., daily, weekly) summarizes all alarm events, identification statistics, and image data to generate structured reports. The correlation analysis module performs deeper analysis, while the spatiotemporal coupling correlation analysis unit analyzes the correlation between alarms at different monitoring points in time and space (e.g., after seepage at a certain point in the dam, does a wet area appear on the downstream slope?). The multiphysics causal inference unit attempts to combine sensor data such as seepage pressure and displacement (if the system is connected) to establish a potential causal relationship model between visual defects and other physical quantities. The risk simulation and early warning unit can perform simple risk extrapolation based on the current defect pattern and evolution trend, and provide preventive maintenance suggestions. The feedback optimization module is the engine for the system's continuous intelligence. The strategy parameter optimization unit collects feedback on the accuracy of scene classification (through manual review or back-calculation of high-confidence recognition results) and the recognition effect after image enhancement. For example, if multiple recognition failures occur in a certain "foggy" scene... This unit analyzes the causes and automatically adjusts the parameter combinations or weights of the contrast enhancement module under the scene classification. The model fine-tuning unit uses high-confidence recognition results accumulated over a period of time (considered as new labeled data) to perform online incremental learning or periodic fine-tuning of the pre-trained model of the deep learning recognition module, enabling the model to better adapt to the subtle features and environmental characteristics of specific water conservancy projects, achieving "increasing accuracy with use." The scheduling optimization module can dynamically adjust the sampling frequency, inspection path, or working mode of cameras at different monitoring points (such as automatically switching to infrared mode at night) based on historical alarm frequencies and scene change patterns, optimizing resource utilization while ensuring monitoring effectiveness. The adaptive update module manages the above optimization process in a coordinated manner, ensuring that the algorithms, parameters, and model versions of all modules can be iteratively updated in an orderly and stable manner, thereby continuously improving the monitoring performance of the entire system over time and with the accumulation of data.

[0024] Example 2 The difference between this embodiment and Embodiment 1 is that the alarm generation module is connected to a data management module. The data management module is responsible for the long-term, structured storage, indexing, and in-depth mining of massive monitoring data (including raw data, enhanced data, identification results, and alarm records). It aims to transform short-term, discrete monitoring events into a traceable, analyzable, and learnable long-term knowledge base, supporting trend prediction and model iteration, and providing high-quality, structured historical data support for model training in the feedback optimization module and causal inference in the correlation analysis module.

[0025] Example 3 The difference between this embodiment and Embodiment 1 is that the data acquisition module is connected to a collaborative monitoring and positioning module. The collaborative monitoring and positioning module is used for wide-area monitoring scenarios such as dams and long-distance rivers. It regards multiple monitoring nodes deployed in different locations (each node includes a camera, sensor and edge computing unit) as intelligent agents, realizes information sharing and collaborative operation among them, so as to complete complex monitoring tasks that a single node cannot handle. It receives the preliminary identification results and raw data from each node, performs fusion and re-analysis, and provides more accurate target positioning and tracking information to the alarm generation module and the correlation analysis module.

[0026] Example 4 The difference between this embodiment and Embodiment 1 is that the deep learning recognition module is connected to a collaborative computing optimization module. The collaborative computing optimization module is used to optimize the distribution of system computing load, balance real-time performance, accuracy and bandwidth cost, and push some lightweight, high real-time processing tasks to the edge side (computing devices close to the camera), while placing complex model inference and global analysis in the cloud to achieve efficient collaboration.

[0027] Example 5 The difference between this embodiment and Embodiment 1 is that the alarm generation module is connected to the human-computer interaction module. The human-computer interaction module is used as a "digital twin operating console" for managers. It deeply integrates the two-dimensional recognition results, alarm information, environmental data and the three-dimensional real scene model (BIM / GIS) of water conservancy facilities, providing an intuitive and immersive monitoring view and interactive analysis tools. This improves the intuitiveness and understandability of the monitoring results, lowers the professional threshold, and assists managers in quickly locating problems, understanding the overall situation, making scientific decisions and emergency command.

[0028] Experimental Example I. Experimental Objective Verify the performance improvement of the dynamic vision-based water conservancy safety monitoring system in the following aspects: Image quality and target visibility in complex environments; The impact of multi-source data fusion on defect detection accuracy; Adaptive optimization and long-term stability; II. Experimental Environment and Dataset Scene: Typical water conservancy facilities such as reservoir dams, floodgates, and dikes; Environmental conditions: sunny day, backlighting, fog and haze, nighttime, rain and snow; Data acquisition equipment: Visible light camera (4K, HDR); Thermal infrared camera (640×512, 25Hz); Environmental sensors (temperature, humidity, illuminance, wind speed); Dataset: 5000 images were collected, and the defect types were labeled, including cracks, seepage, deformation, vegetation invasion, etc. III. Comparison Methods Traditional methods: Visual monitoring systems based on fixed threshold segmentation and edge detection; Multi-sensor approach: Visible light + infrared independent analysis, results then fused; This system is a fusion and adaptive enhancement system based on dynamic vision. IV. Evaluation Indicators Image quality: PSNR, SSIM, image information entropy; Recognition accuracy: precision, recall, F1 score; Environmental adaptability: performance retention rate under different lighting / weather conditions; System response time: The average time from data acquisition to alarm generation; V. Experimental Data ; ; VIII. Experimental Conclusions The table above shows the environmental adaptability: Under adverse conditions such as backlight, haze, and nighttime, the image quality and recognition accuracy of this system are significantly better than traditional methods and multi-sensor methods, especially in low-light environments where the PSNR is improved by about 29%; Multi-source fusion advantages: Through feature alignment and fusion mechanisms, the system improves the defect detection F1 score to 92.8% with the complementarity of infrared and visible light data, and still maintains a high recognition capability when a single sensor fails; Adaptability and stability: The system exhibits stronger stability during long-term operation (F1 fluctuation is only ±2.1%), indicating that the feedback optimization and dynamic scheduling mechanism is effective; Real-time performance: The average processing latency is reduced to 680ms, meeting the near real-time requirements of water conservancy safety monitoring.

[0029] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A water conservancy safety monitoring system based on dynamic vision, characterized in that: include: The data acquisition module is connected to a lighting evaluation module, which is connected to a scene classification module, which is connected to a region segmentation module, and the region segmentation module is connected to a contrast enhancement module. A quality assessment module is connected to a contrast enhancement module, a dynamic adjustment module is connected to a fusion optimization module, and a feature alignment and fusion module is connected to a feature alignment and fusion module. A deep learning recognition module is connected to a fusion optimization module and a feature alignment and fusion module. The deep learning recognition module is also connected to a comparison analysis module, which is in turn connected to a reconstruction module. An alarm generation module is connected to a comparison analysis module and a deep learning recognition module. The alarm generation module is also connected to a report generation module, which is connected to a correlation analysis module. The feedback optimization module is connected to the scene classification module and the alarm generation module. The feedback optimization module is also connected to the scheduling optimization module, which in turn is connected to the adaptive update module.

2. The water conservancy safety monitoring system based on dynamic vision according to claim 1, characterized in that: The data acquisition module includes a multispectral image acquisition unit, a high dynamic range imaging unit, and an environmental parameter synchronous sensing unit. The multispectral image acquisition unit is connected to the high dynamic range imaging unit, and the high dynamic range imaging unit is connected to the environmental parameter synchronous sensing unit. The data acquisition module is responsible for the synchronous and high-quality acquisition of raw monitoring data, providing a multi-dimensional information source for subsequent processing.

3. The water conservancy safety monitoring system based on dynamic vision according to claim 2, characterized in that: The scene classification module includes an illumination pattern discrimination unit, a weather interference assessment unit, and a scene semantic preliminary judgment unit. The illumination pattern discrimination unit is connected to the weather interference assessment unit, and the weather interference assessment unit is connected to the scene semantic preliminary judgment unit. The scene classification module is used to intelligently identify and classify the current monitoring scene to determine the basic strategy for subsequent image processing.

4. A water conservancy safety monitoring system based on dynamic vision according to claim 3, characterized in that: The region segmentation module includes a region of interest generation unit, a texture analysis unit, and an edge optimization unit. The region of interest generation unit is connected to the texture analysis unit, and the texture analysis unit is connected to the edge optimization unit. The region segmentation module is used to perform refined and structured region division on the monitoring image.

5. A water conservancy safety monitoring system based on dynamic vision according to claim 4, characterized in that: The contrast enhancement module includes a spatial enhancement unit, a frequency domain illumination correction unit, and a mapping fusion unit. The spatial enhancement unit is connected to the frequency domain illumination correction unit, and the frequency domain illumination correction unit is connected to the mapping fusion unit. The contrast enhancement module is used to maximize the visual separability of target features and background in response to the characteristics of different regions and the problem of insufficient contrast. The spatial enhancement unit enhances contrast by adjusting the pixel values ​​of the image, using a histogram equalization algorithm; Histogram equalization is a process that adjusts the histogram distribution of an image to make the pixel values ​​more evenly distributed. The steps include: Calculate the histogram of the image ,in Represents the pixel grayscale value. Indicates grayscale value The number of pixels; Calculate the cumulative histogram : ； The cumulative histogram is normalized to the range [0, 255] to obtain new grayscale values. : ； in, The total number of pixels in the image; The grayscale values ​​in the original image are... The pixels are replaced with new grayscale values. This yields an image with enhanced contrast. The frequency domain illumination correction unit processes the image in the frequency domain to correct the effect of uneven illumination on image contrast, using a high-pass filtering algorithm. High-pass filtering algorithms enhance image contrast by filtering out low-frequency components and retaining high-frequency components. The steps include: To transform an image from the spatial domain to the frequency domain, the Fast Fourier Transform is typically used. ； in, These are the pixel values ​​of the image. It is its frequency domain representation; Design a high-pass filter. A high-pass filter is an ideal high-pass filter, defined as follows: ； in, It is a frequency point Distance to the center of the frequency domain It is the cutoff frequency; Multiplying the filter by the frequency domain representation of the image yields the filtered frequency domain image: ； To convert the filtered frequency domain image back to the spatial domain, the inverse fast Fourier transform is typically used. ； in, It is an image with enhanced contrast; The mapping fusion unit fuses the images after spatial domain enhancement and frequency domain illumination correction to obtain better contrast enhancement effect, using a weighted average fusion algorithm; The weighted average fusion algorithm fuses multiple images by adding them together according to certain weights. The steps include: Let the spatially enhanced image be The image after frequency domain illumination correction is ; Choose appropriate weights and ,satisfy ; Calculate the fused image ; ； By adjusting the weights and This allows for the control of the impact of spatial domain enhancement and frequency domain illumination correction on the final image, thereby achieving the best contrast enhancement effect.

6. A water conservancy safety monitoring system based on dynamic vision according to claim 5, characterized in that: The feature alignment and fusion module includes a spatiotemporal registration unit, a feature-level fusion unit, and a saliency enhancement unit. The spatiotemporal registration unit is connected to the feature-level fusion unit, and the feature-level fusion unit is connected to the saliency enhancement unit. The feature alignment and fusion module is used to achieve the unification and complementarity of multi-source heterogeneous data at the spatiotemporal and feature levels, providing more comprehensive information for identification.

7. A water conservancy safety monitoring system based on dynamic vision according to claim 6, characterized in that: The deep learning recognition module includes a multi-task recognition unit, an enhanced detection unit, and an inference unit. The multi-task recognition unit is connected to the enhanced detection unit, and the enhanced detection unit is connected to the inference unit. The deep learning recognition module is used to automatically, accurately, and in parallel detect and identify various typical defects and anomalies of water conservancy projects on images that integrate multi-dimensional features, based on a trained deep neural network model.

8. A water conservancy safety monitoring system based on dynamic vision according to claim 7, characterized in that: The alarm generation module includes a multi-level alarm rule unit, an alarm evaluation unit, and an alarm information generation unit. The multi-level alarm rule unit is connected to the alarm evaluation unit, and the alarm evaluation unit is connected to the alarm information generation unit. The alarm generation module is used to transform the analysis and identification results into standardized and operable early warning information and trigger subsequent processes.

9. A water conservancy safety monitoring system based on dynamic vision according to claim 8, characterized in that: The feedback optimization module includes a strategy parameter optimization unit, a dynamic scheduling unit, and a model fine-tuning unit. The strategy parameter optimization unit is connected to the dynamic scheduling unit, and the dynamic scheduling unit is connected to the model fine-tuning unit. The feedback optimization module is used to realize the system's self-learning and continuous improvement.

10. A water conservancy safety monitoring system based on dynamic vision according to claim 9, characterized in that: The correlation analysis module includes a spatiotemporal coupling correlation analysis unit, a multiphysics causal inference unit, and a risk simulation and early warning unit. The spatiotemporal coupling correlation analysis unit is connected to the multiphysics causal inference unit, and the multiphysics causal inference unit is connected to the risk simulation and early warning unit. The correlation analysis module is used to explore the intrinsic relationship between different alarm events and monitoring parameters.

Images