Deep Learning-Based Methods, Systems, and Equipment for Identifying Seepage Points in Flood Control Embankments

Abstract

This application provides a method, system, and device for identifying seepage points in flood control dikes based on deep learning. The method includes: acquiring visible light and infrared images of the dike area to be detected; extracting the main body area of ​​the dike based on the visible light image; inputting the infrared image corresponding to the main body area of ​​the dike into a deep learning model, and identifying candidate seepage areas and their seepage confidence levels through the deep learning model; acquiring the temperature difference features of the candidate seepage areas in the infrared image; determining effective seepage points based on the temperature difference features, geometric features in the visible light image, and seepage confidence levels of the candidate seepage areas; outputting the location information of the effective seepage points based on the registration relationship between the visible light and infrared images; fusing the information from the visible light and infrared images, first accurately extracting the main body area of ​​the dike based on the visible light image, and then combining the deep learning model to identify candidates for infrared thermal anomalies, thereby improving the accuracy and reliability of dike seepage identification.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a method, system and device for identifying seepage points in flood control dikes based on deep learning. Background Technology

[0002] During long-term service, flood control dikes are susceptible to factors such as water pressure, rainwater erosion, foundation settlement, and material aging, leading to potential hazards such as seepage, cracks, and piping. Traditional manual inspection methods are limited by experience and inspection frequency, making it difficult to detect early, hidden seepage in a timely manner. While sensor monitoring offers a degree of continuity, its deployment cost is high and its coverage is limited. Infrared thermal imaging technology can detect potential seepage areas through surface temperature anomalies, offering advantages such as non-contact and rapid inspection. However, existing methods often rely on human experience or simple threshold judgments, making them susceptible to changes in ambient temperature, solar radiation, vegetation obstruction, and shooting angle, resulting in false alarms, missed alarms, and inaccurate location. Therefore, there is an urgent need for a method that can combine deep learning algorithms to achieve intelligent identification of dike seepage points, thereby improving identification accuracy, automation level, and engineering applicability. Summary of the Invention

[0003] In view of this, it is necessary to provide a method, system and equipment for identifying seepage points in flood control dikes based on deep learning, which can at least overcome one of the above-mentioned defects.

[0004] In a first aspect, embodiments of this application provide a method for identifying seepage points in flood control dikes based on deep learning, the method comprising:

[0005] Acquire visible light and infrared images of the dam area to be inspected;

[0006] Extract the main area of ​​the dam based on the visible light image;

[0007] The infrared image corresponding to the main area of ​​the dam is input into a deep learning model, and the deep learning model is used to identify seepage candidate areas and the seepage confidence of the seepage candidate areas.

[0008] Obtain the temperature difference characteristics of the candidate leakage area in the infrared image;

[0009] Valid leakage points are determined based on the temperature difference characteristics of the candidate leakage areas, their geometric morphological characteristics in visible light images, and the leakage confidence level.

[0010] Based on the registration relationship between the visible light image and the infrared image, the location information of the effective leakage point is output.

[0011] In one embodiment, determining the effective leak point based on the temperature difference characteristics of the candidate leak region, its geometric morphological features in a visible light image, and the leak confidence level includes:

[0012] Based on the location of the leakage candidate region in the visible light image, the geometric morphological features are extracted, including the area of ​​the connected region, the aspect ratio, and the integrity of the edge contour.

[0013] The temperature difference characteristics, the geometric morphology characteristics, and the leakage confidence level are weighted and calculated to obtain a comprehensive evaluation score.

[0014] If the comprehensive evaluation score is greater than the preset score threshold, the candidate leakage area is determined to be a valid leakage point.

[0015] In one embodiment, obtaining the temperature difference characteristics of the leakage candidate region in the infrared image includes:

[0016] Extract the first average temperature of the candidate leakage region in the infrared image;

[0017] The background neighborhood of the candidate leakage area is determined after expanding outward by a preset range in the infrared image, and the second average temperature of the background neighborhood is extracted;

[0018] The difference between the second average temperature and the first average temperature is used as the temperature difference characteristic of the leakage candidate area.

[0019] In one embodiment, the deep learning model integrates an attention enhancement mechanism;

[0020] The process of identifying leakage candidate regions using the deep learning model includes:

[0021] The spatial attention module in the deep learning model is applied to enhance the weights of pixels with significant temperature differences in the main region of the dam.

[0022] The channel attention module is used to extract thermal anomaly features at different scales from the infrared image to identify leakage morphology features.

[0023] In one embodiment, the method further includes:

[0024] Based on the temperature difference characteristic intensity, area range, and spatially sensitive location of the effective seepage point in the main area of ​​the dam, the effective seepage point is classified into risk levels.

[0025] Based on the risk level, a corresponding graded early warning signal is output, and the graded early warning signal is associated with the location information and stored in the disease database of the dam area to be detected.

[0026] In one embodiment, determining the effective leak point based on the temperature difference characteristics of the candidate leak region, its geometric morphological features in a visible light image, and the leak confidence level includes:

[0027] The environmental parameters during the infrared image acquisition are obtained, including sunlight intensity, wind speed, and ambient temperature.

[0028] An environmental correction factor is determined based on the environmental parameters, and the leakage confidence level is dynamically corrected using the environmental correction factor.

[0029] In one embodiment, the step of dynamically correcting the leakage confidence level using the environmental correction coefficient includes:

[0030] Multiply the leakage confidence level by the environmental correction factor to obtain the corrected leakage score;

[0031] The candidate leakage regions are reassessed based on the corrected leakage score to suppress false thermal anomalies.

[0032] In one embodiment, the method further includes:

[0033] Leakage identification is performed on multiple consecutively acquired infrared images to obtain the confidence sequence of the same leakage candidate region at different time points;

[0034] The confidence sequence is smoothed to calculate the time-series cumulative confidence of the leakage candidate region;

[0035] If the time-series cumulative confidence of the leakage candidate region shows an increasing trend or remains stable, then the confidence of the leakage candidate region is increased.

[0036] Secondly, this application provides a deep learning-based system for identifying seepage points in flood control dikes, used to implement the method described in the first aspect, the system comprising:

[0037] The information acquisition module is used to acquire visible light and infrared images of the dam area to be detected;

[0038] The feature extraction module is used to extract the main body area of ​​the dam based on the visible light image, input the infrared image corresponding to the main body area of ​​the dam into a deep learning model, identify the leakage candidate area and the leakage confidence of the leakage candidate area through the deep learning model, and obtain the temperature difference feature of the leakage candidate area in the infrared image.

[0039] The result output module is used to determine the effective leakage point based on the temperature difference characteristics of the leakage candidate area, the geometric morphology characteristics in the visible light image, and the leakage confidence level; and to output the location information of the effective leakage point based on the registration relationship between the visible light image and the infrared image.

[0040] Thirdly, embodiments of this application provide an electronic device, including:

[0041] Processor; and

[0042] A memory having computer-readable instructions stored thereon for controlling the processor to perform the method as described in the first aspect.

[0043] This application provides a method, system, and device for identifying seepage points in flood control dikes based on deep learning. By fusing visible light and infrared image information, the main area of ​​the dike is first accurately extracted based on the visible light image. Then, the infrared thermal anomaly is candidate-based by combining a deep learning model. Temperature difference features and geometric morphology features are introduced for secondary judgment, forming a two-stage discrimination mechanism of "model recognition + rule verification". This significantly improves the ability to identify early and weak seepage anomalies and reduces false alarms and missed alarms caused by environmental interference. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a method for identifying seepage points in flood control dikes based on deep learning, provided in one embodiment of this application.

[0045] Figure 2 This is a schematic diagram of a module of a deep learning-based flood control dike seepage point identification system provided in one embodiment of this application.

[0046] Figure 3 This is a schematic diagram of the modules of an electronic device provided in an embodiment of this application.

[0047] Explanation of main component symbols

[0048] The system for identifying seepage points in flood control dikes based on deep learning consists of: 10, information acquisition module 11, feature extraction module 12, result output module 13, electronic device 20, processor 21 and memory 22, and method steps S100-S600. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0050] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0051] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0052] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] During long-term service, flood control dikes are susceptible to factors such as water pressure, rainwater erosion, foundation settlement, and material aging, leading to potential hazards such as seepage, cracks, and piping. Traditional manual inspection methods are limited by experience and inspection frequency, making it difficult to detect early, hidden seepage in a timely manner. While sensor monitoring offers a degree of continuity, its deployment cost is high and its coverage is limited. Infrared thermal imaging technology can detect potential seepage areas through surface temperature anomalies, offering advantages such as non-contact and rapid inspection. However, existing methods often rely on human experience or simple threshold judgments, making them susceptible to changes in ambient temperature, solar radiation, vegetation obstruction, and shooting angle, resulting in false alarms, missed alarms, and inaccurate location. Therefore, there is an urgent need for a method that can combine deep learning algorithms to achieve intelligent identification of dike seepage points, thereby improving identification accuracy, automation level, and engineering applicability.

[0054] In view of this, the deep learning-based method, system and equipment for identifying seepage points in flood control dikes provided in this application, by fusing visible light image and infrared image information, first accurately extracts the main area of ​​the dike based on the visible light image, then combines the deep learning model to identify candidates for infrared thermal anomalies, and introduces temperature difference features and geometric morphology features for secondary judgment, forming a two-stage discrimination mechanism of "model recognition + rule verification", thereby significantly improving the ability to identify early weak seepage anomalies and reducing false alarms and missed alarms caused by environmental interference.

[0055] Meanwhile, this application establishes a registration relationship between infrared and visible light images to achieve precise location of seepage points in actual dam scenarios. Combined with environmental correction coefficients and time-series confidence accumulation mechanisms, it effectively suppresses false thermal anomalies. On this basis, a risk level classification and early warning linkage mechanism is introduced so that the identification results can directly serve dam inspection and disease management, thereby improving the intelligence and engineering practicality of flood control dam safety monitoring.

[0056] Figure 1 This is a flowchart illustrating a deep learning-based method for identifying seepage points in flood control dikes, as provided in one embodiment of this application. Figure 1 The deep learning-based method for identifying seepage points in flood control dikes includes at least the following steps: S100: acquiring visible light and infrared images of the dike area to be detected; S200: extracting the main body area of ​​the dike based on the visible light image; S300: inputting the infrared image corresponding to the main body area of ​​the dike into a deep learning model, and identifying candidate seepage areas and their seepage confidence levels through the deep learning model; S400: acquiring the temperature difference features of the candidate seepage areas in the infrared image; S500: determining effective seepage points based on the temperature difference features, geometric features in the visible light image, and seepage confidence levels of the candidate seepage areas; S600: outputting the location information of the effective seepage points based on the registration relationship between the visible light image and the infrared image.

[0057] S100: Acquire visible light and infrared images of the dam area to be detected.

[0058] In this embodiment of the application, the deep learning-based flood control dike seepage point identification method includes, in step S100, acquiring visible light images and infrared images of the dike area to be detected.

[0059] Specifically, drones, vehicle-mounted inspection equipment, handheld inspection terminals, or fixed monitoring devices can be used to acquire images of the dam area to be inspected. The acquired visible light images are used to characterize the geometric structure, texture contours, and target boundaries of the dam surface, while the acquired infrared images are used to characterize the temperature distribution and thermal anomaly characteristics of the dam surface. Optionally, environmental parameters such as acquisition time, ambient temperature, wind speed, and solar radiation intensity can also be acquired simultaneously to facilitate subsequent correction of the infrared image recognition results.

[0060] It is understandable that seepage points in dams are usually not obvious in visible light images, but may appear as temperature anomalies in infrared images due to local water content changes. Therefore, acquiring both visible light and infrared images simultaneously helps to take into account both the spatial structure information and thermal anomaly information of the dam area, providing a more comprehensive data foundation for subsequent identification.

[0061] S200: Extracting the main area of ​​the dam based on visible light images.

[0062] In this embodiment of the application, the method for identifying seepage points in flood control dikes based on deep learning includes extracting the main area of ​​the dike based on visible light images in step S200.

[0063] Specifically, the acquired visible light image can be input into a target detection model, semantic segmentation model, or edge extraction model to automatically locate and segment the main area of ​​the dam, thereby separating the target areas such as the dam body, slope, shoulders, and toe from the background areas such as the sky, water surface, vegetation, roads, and buildings. Furthermore, a region mask can be generated from the extracted main dam area and used for region matching and feature constraint of subsequent infrared images.

[0064] Understandably, the environment surrounding a dam often contains numerous distracting targets unrelated to seepage. Directly identifying the entire image could easily introduce background noise and reduce the model's accuracy. Therefore, first extracting the main dam area from the visible light image effectively narrows the identification range, reduces the impact of irrelevant background on seepage identification results, and improves the accuracy and stability of subsequent feature extraction.

[0065] S300: Input the infrared image corresponding to the main area of ​​the dam into the deep learning model, and identify the seepage candidate area and the seepage confidence of the seepage candidate area through the deep learning model.

[0066] In this embodiment of the application, the method for identifying seepage points in flood control dikes based on deep learning includes step S300, which involves inputting an infrared image corresponding to the main area of ​​the dike into a deep learning model, and identifying candidate seepage areas and the seepage confidence level of the candidate seepage areas through the deep learning model.

[0067] Specifically, the deep learning-based flood control dike seepage point identification method includes step S300, which involves inputting the infrared image corresponding to the main area of ​​the dike into a deep learning model, and identifying seepage candidate areas and the seepage confidence of the seepage candidate areas through the deep learning model.

[0068] Specifically, based on the dam's main body region obtained in step S200, the infrared image can be cropped or spatially mapped, so that the portion of the infrared image corresponding to the dam's main body region is input into a deep learning model. The deep learning model can employ a convolutional neural network, encoder-decoder network, attention enhancement network, or multi-scale feature fusion network to identify local thermal anomalies in the infrared image and output one or more leakage candidate regions and their corresponding leakage confidence scores. The leakage confidence score characterizes the probability that the candidate region is a leakage region.

[0069] Understandably, seepage points in flood control dikes often appear in infrared images as areas of localized low temperature, blurred boundaries, irregular shapes, or banded diffusion of thermal anomalies, making accurate identification difficult solely based on human experience. Automated analysis of infrared images using deep learning models can extract high-level semantic features related to seepage from complex backgrounds, thereby improving the ability to identify early, subtle seepage phenomena.

[0070] S400: Obtain the temperature difference characteristics of the leakage candidate area in the infrared image.

[0071] In this embodiment of the application, the deep learning-based flood control dike seepage point identification method includes, in step S400, obtaining the temperature difference features of the seepage candidate area in the infrared image.

[0072] Specifically, for each candidate leakage area identified in step S300, the average temperature, minimum temperature, maximum temperature, and temperature dispersion of the candidate area can be statistically analyzed in the infrared image. Furthermore, a preset range is expanded around the candidate area to form a background neighborhood, from which the corresponding average temperature or local reference temperature is extracted. Subsequently, the temperature difference between the candidate leakage area and its background neighborhood is used as a temperature difference feature to characterize whether the area has a significant thermal anomaly.

[0073] Understandably, dam seepage is often accompanied by increased local water content and changes in thermal conductivity, resulting in temperature differences between the affected area and the surrounding normal area. By calculating temperature difference features, the candidate regions output by the deep learning model can be further quantified, allowing subsequent judgments to not only rely on model confidence but also incorporate more physically meaningful thermal anomaly information, thereby improving the reliability of the identification results.

[0074] S500: Determine the effective leakage point based on the temperature difference characteristics of the leakage candidate area, the geometric morphology characteristics in the visible light image, and the leakage confidence level.

[0075] In this embodiment of the application, the deep learning-based flood control dike seepage point identification method includes, in step S500, determining the effective seepage point based on the temperature difference characteristics of the seepage candidate area, the geometric morphology characteristics in the visible light image, and the seepage confidence level.

[0076] Specifically, geometric features can be extracted based on the corresponding location of the candidate leakage area in the visible light image. These geometric features may include the area of ​​connected regions, aspect ratio, edge contour integrity, shape regularity, or consistency of diffusion direction. Subsequently, the temperature difference features, geometric features, and leakage confidence are weighted and fused to obtain a comprehensive evaluation result. When the comprehensive evaluation result meets the preset judgment conditions, the corresponding candidate area is determined as a valid leakage point.

[0077] Understandably, relying solely on infrared thermal anomalies is susceptible to interference from factors such as sunlight, shadows, and differences in surface materials, while relying solely on visible light geometric features is insufficient to reflect the true thermal response of leaks. Therefore, combining temperature difference features, geometric morphology features, and leak confidence levels for joint judgment can simultaneously verify the authenticity of candidate regions from three levels: thermal anomalies, spatial structure, and model semantics. This effectively reduces the false alarm rate and improves the accuracy of screening for true leak points.

[0078] S600: Based on the registration relationship between the visible light image and the infrared image, output the location information of the effective leakage point.

[0079] In this embodiment of the application, the deep learning-based flood control dike seepage point identification method includes, in step S600, outputting the location information of the effective seepage point according to the registration relationship between the visible light image and the infrared image.

[0080] Specifically, a spatial registration relationship between visible light and infrared images can be established in advance, mapping the pixel coordinates of effective seepage points in the infrared image to their corresponding positions in the visible light image. This is further combined with the geographical coordinates, mileage information, or inspection trajectory information of the dam to determine the spatial location of the effective seepage points in the actual dam scenario. Optionally, the output results can be annotated on the visible light image to generate a seepage point identification result map with location markers, facilitating rapid location and subsequent handling by inspection personnel.

[0081] Understandably, infrared images primarily reflect temperature information, while visible light images are better suited for representing the spatial morphology and actual location of a target. By establishing a registration relationship between the two types of images, not only can the location of the leak point in the image be accurately output, but it can also be further mapped to the actual location of the dam, thereby enhancing the engineering usability of the detection results and facilitating on-site personnel to conduct verification, maintenance, and risk assessment.

[0082] The deep learning-based method for identifying seepage points in flood control dikes provided in this application can automatically extract candidate seepage areas, analyze temperature anomalies, determine geometric features, and output precise locations. Compared to traditional methods that rely on manual inspections or single threshold judgments, this solution effectively improves the accuracy, stability, and automation level of seepage point identification, reduces false alarms and missed alarms caused by environmental interference, and provides reliable data support for intelligent inspection, risk warning, and defect treatment of flood control dikes.

[0083] In this embodiment, determining a valid leak point based on the temperature difference characteristics of the leak candidate area, its geometric morphological features in the visible light image, and the leak confidence score includes: extracting geometric morphological features based on the corresponding position of the leak candidate area in the visible light image, the geometric morphological features including the area of ​​connected regions, aspect ratio, and edge contour integrity; weighting the temperature difference features, geometric morphological features, and leak confidence score to obtain a comprehensive evaluation score; if the comprehensive evaluation score is greater than a preset score threshold, the leak candidate area is determined to be a valid leak point.

[0084] Specifically, let the temperature difference characteristics of the leakage candidate region be: Geometric morphological features are The leakage confidence level is The comprehensive evaluation score can then be expressed as:

[0085]

[0086] in, To comprehensively evaluate the score, , and These are the weighting coefficients corresponding to temperature difference characteristics, geometric morphology characteristics, and leakage confidence levels, respectively, and satisfy the following conditions: ; , and These represent the normalized temperature difference characteristic value, geometric morphology characteristic value, and leakage confidence value, respectively. Furthermore, when... When the leakage candidate area is determined as the effective leakage point, then, This is a preset score threshold.

[0087] Understandably, determining whether a candidate leakage area truly constitutes a leakage point cannot rely solely on a single temperature anomaly or a single model output. By jointly weighting temperature difference characteristics, geometric morphology characteristics, and leakage confidence, a comprehensive screening can be conducted from three dimensions: thermal response intensity, reasonableness of the area's morphology, and the reliability of the model's judgment. This effectively improves the accuracy of leakage point identification and reduces misjudgments caused by background interference, local thermal noise, or occasional anomalies.

[0088] In this embodiment of the application, obtaining the temperature difference feature of the leakage candidate area in the infrared image includes: extracting the first average temperature of the leakage candidate area in the infrared image; determining the background neighborhood of the leakage candidate area after expanding outward by a preset range in the infrared image, and extracting the second average temperature of the background neighborhood; and using the difference between the second average temperature and the first average temperature as the temperature difference feature of the leakage candidate area.

[0089] Specifically, let the set of pixels within the candidate leakage region be . The set of background neighborhood pixels is Pixels in infrared images Temperature value Then the first average temperature With the second average temperature They can be represented as:

[0090]

[0091]

[0092] in, and These represent the number of pixels in the candidate leakage region and the background neighborhood, respectively. The temperature difference feature... It can be represented as:

[0093]

[0094] When the candidate leakage area exhibits low temperature anomalies The value is usually positive, and the larger the value, the more significant the temperature difference between the candidate region and the surrounding background.

[0095] It is understandable that when a flood control embankment leaks, the increased local water content alters the material's thermal conductivity and heat capacity properties, resulting in a different thermal distribution in the infrared image compared to the background area. By comparing the average temperature difference between a candidate region and its neighborhood, the degree of thermal anomaly in the candidate region can be quantified, making subsequent judgments more physically meaningful and reducing the instability caused by relying solely on image texture.

[0096] In this embodiment, the deep learning model integrates an attention enhancement mechanism; the identification of seepage candidate areas through the deep learning model includes: applying the spatial attention module in the deep learning model to enhance the weight of pixels with significant temperature differences in the main area of ​​the dam; and applying the channel attention module to extract thermal anomaly features at different scales in the infrared image to identify seepage morphology features.

[0097] Specifically, the spatial attention module can assign weights to spatial locations in the infrared image, resulting in higher response values ​​for areas with significant local temperature differences associated with leakage, while the weight of background areas is relatively suppressed. The channel attention module can adaptively recalibrate multi-channel feature maps, highlighting thermal textures, edge variations, and local diffusion patterns associated with leakage at different scales. Optionally, the attention mechanism can be applied to the encoding layer, feature fusion layer, or decoding layer of the deep learning model to enhance the network's ability to identify small leaks, leaks with blurred edges, and leaks with low contrast.

[0098] Understandably, dam seepage in infrared images typically exhibits weak features, strong localization, and significant background interference. Without an attention mechanism, the model might easily misinterpret non-seepage factors such as solar radiation, a wetted surface, and shadow changes as thermal anomalies. By combining spatial and channel attention, the model can better focus on key areas and features related to seepage, thereby improving the deep learning model's ability and robustness in identifying seepage candidate areas in complex scenarios.

[0099] In this embodiment, the method further includes: classifying the effective seepage points into risk levels based on their temperature difference characteristics, area range, and spatially sensitive location within the main dam area; outputting corresponding graded early warning signals based on the risk levels; and storing the graded early warning signals in a disease database of the dam area to be inspected, in association with the location information.

[0100] Specifically, it can be based on the temperature difference characteristics of the effective leakage points. ,area and spatial sensitivity coefficient A risk score is constructed, whereby the spatial sensitivity coefficient characterizes the risk weighting when the leakage point is located in critical locations such as the dam toe, dam shoulder, upstream slope, joint area, or weak drainage zone. The risk score can be expressed as:

[0101]

[0102] in, To score the risk, , , The corresponding weight coefficients, and satisfying ; , and These represent the normalized temperature difference intensity, area range, and spatial sensitivity coefficient, respectively. According to The size of the leak can be used to classify effective leakage points into different risk levels, such as low risk, medium risk, and high risk, and output corresponding early warning levels for each. Furthermore, the risk level, location information, time information, and image identification information are all written into the disease database for subsequent tracking analysis, trend judgment, and maintenance.

[0103] Understandably, not all seepage points exhibit the same level of severity. Generally, the greater the temperature difference, the larger the area affected, and the closer the location is to critical stress or weak points in the dam, the higher the likelihood of it evolving into serious damage. By constructing a risk score and implementing tiered early warning systems, the identification results can directly serve dam operation and maintenance management, helping inspection personnel prioritize high-risk areas and improving the timeliness and targeted nature of damage treatment.

[0104] This application provides a deep learning-based method for identifying seepage points in flood control dikes. This method combines infrared thermal anomaly features, visible light geometric features, and deep learning confidence levels to comprehensively determine candidate seepage areas from multiple dimensions. Furthermore, it improves identification accuracy and early warning reliability through attention enhancement, weighted evaluation, and risk grading mechanisms. Compared to existing methods that rely solely on a single infrared threshold or manual experience, this solution effectively reduces false alarms and false negatives, enhances the ability to detect early, hidden seepage, and achieves accurate identification, precise location, risk grading, and database recording of seepage points. It possesses strong engineering practicality and widespread application value.

[0105] In this embodiment of the application, determining the effective leakage point based on the temperature difference characteristics of the leakage candidate area, the geometric morphological characteristics in the visible light image, and the leakage confidence level includes: acquiring environmental parameters during infrared image acquisition, including solar radiation intensity, wind speed, and ambient temperature; determining an environmental correction coefficient based on the environmental parameters; and applying the environmental correction coefficient to dynamically correct the leakage confidence level.

[0106] Specifically, solar radiation intensity, wind speed, and ambient temperature can be denoted as follows: , and The environmental correction coefficient is determined based on a preset mapping relationship or an empirical model. For example, the environmental correction factor can be expressed as:

[0107]

[0108] in, , and These are the correction functions corresponding to solar radiation intensity, wind speed, and ambient temperature, respectively. , , Let be the weight coefficient, and satisfy... Furthermore, the functions can be set to monotonic mapping according to the data collection scenario, so that the environmental correction coefficient is reduced accordingly under conditions of strong sunlight, high wind speed, or large ambient temperature difference, thereby reducing the judgment weight of the candidate region.

[0109] It is understandable that thermal anomalies in dam infrared images are not entirely caused by seepage; external environmental factors also affect the temperature distribution in the images. For example, strong sunlight may cause localized surface heating, high wind speeds may accelerate surface heat dissipation, and changes in ambient temperature can cause overall thermal field drift. By introducing environmental parameters and constructing environmental correction coefficients, the seepage confidence level can be adaptively corrected, thereby reducing the impact of external interference on the identification results and improving the reliability of valid seepage point determination.

[0110] In this embodiment of the application, the leakage confidence is dynamically corrected by applying an environmental correction coefficient, including: multiplying the leakage confidence by the environmental correction coefficient to obtain a corrected leakage score; and re-evaluating the leakage candidate area based on the corrected leakage score to suppress false thermal anomalies.

[0111] Specifically, let the original leakage confidence level of the deep learning model output be... The environmental correction factor is The corrected leakage score is... It can be represented as:

[0112]

[0113] in, Used to characterize the initial probability that a candidate leakage area is a real leakage point. Used to characterize the degree of correction of this probability under the current environmental conditions. This score is used to represent the final judgment score after environmental impact correction. Furthermore, it can be used to determine the leakage score after correction. Compared with a preset threshold, if If so, then retain that region as a valid candidate region; if If it is, then its judgment priority will be reduced or it will be directly removed.

[0114] Understandably, while deep learning models can identify thermal anomalies in infrared images, some of these anomalies may originate from environmental disturbances rather than actual leaks. By dynamically correcting the leak confidence level, the model's output can better reflect real-world inspection scenarios, avoiding misidentification of false hot spots caused by external factors as leak points, thereby improving the stability and accuracy of the identification results.

[0115] In this embodiment of the application, the method further includes: performing leakage identification on multiple continuously acquired infrared images to obtain a confidence sequence of the same leakage candidate region at different time nodes; smoothing the confidence sequence to calculate the temporal cumulative confidence of the leakage candidate region; and increasing the confidence of the leakage candidate region if the temporal cumulative confidence of the leakage candidate region shows an increasing trend or remains stable.

[0116] Specifically, let the same candidate leakage area be in the th... The confidence level in the frame infrared image is Then its confidence sequence can be constructed. Furthermore, the sequence can be smoothed using moving averages or exponential smoothing. For example, when using exponential smoothing, the cumulative confidence level of the time series... It can be represented as:

[0117]

[0118] in, For the first The temporal cumulative confidence level corresponding to the frame. The temporal cumulative confidence score for the previous frame. It is the smoothing coefficient, and In multiple consecutive frames If the temperature shows a continuous upward trend or remains at a high level at multiple time points, the candidate region can be considered to have strong persistent thermal anomaly characteristics, thereby increasing its final leakage confidence.

[0119] Understandably, genuine leaks are typically persistent and gradual, not merely appearing briefly in a single frame. False thermal anomalies, caused by factors such as rainwater reflection, surface shadows, and localized thermal disturbances, are often sporadic and unstable. By performing temporal analysis and confidence accumulation on multiple infrared images, the ability to identify persistent leaks can be enhanced, while suppressing transient noise interference and increasing the probability of detecting early, concealed leaks.

[0120] This application's embodiments introduce an environmental parameter correction mechanism and a multi-frame temporal accumulation mechanism to perform secondary correction and dynamic optimization on the leakage confidence score output by the deep learning model. This effectively reduces the impact of solar radiation, wind speed, ambient temperature fluctuations, and instantaneous thermal noise on the leakage identification results, and improves the ability to suppress false thermal anomalies and the stability of identifying real leakage points. At the same time, by combining the aforementioned temperature difference features, geometric morphology features, and confidence score fusion judgment method, the accuracy, robustness, and engineering applicability of flood control dike leakage point identification can be further improved.

[0121] The following exemplary embodiment describes the complete process of the deep learning-based method for identifying seepage points in flood control dikes provided in this application. Assume the object to be detected is a section of flood control dike, which requires regular inspections before and after the flood season. Inspection personnel use inspection equipment equipped with a visible light camera and an infrared thermal imager to simultaneously collect images of the dike's upstream slope, downstream slope, toe, and joint areas, obtaining visible light and infrared images of the dike area to be detected.

[0122] In this embodiment, visible light and infrared images of the dam area to be inspected are first acquired, and environmental parameters at the time of acquisition are recorded, including solar radiation intensity, wind speed, and ambient temperature. Specifically, the inspection equipment takes pictures of the dam on a sunny morning. At this time, the visible light image can clearly show the outline of the dam, slope structure, vegetation distribution, and local water accumulation traces, while the infrared image reflects the temperature distribution on the dam surface. It is understood that the geometric location of the same position is usually easier to identify in the visible light image, while local temperature differences caused by seepage are easier to detect in the infrared image. Therefore, the simultaneous acquisition of dual-modal images can provide a more comprehensive data foundation for subsequent identification.

[0123] In this embodiment, the main dam region is extracted based on the visible light image. Specifically, the visible light image can be input into a pre-trained semantic segmentation model to automatically segment the dam body, slope, and toe areas, and remove non-target areas such as the sky, water surface, trees, roads, and building shadows to obtain a mask of the main dam region. Further, this mask is mapped onto an infrared image to obtain the infrared image region corresponding to the main dam region. It is understood that the surrounding environment of a dam often contains many interfering targets. If the main region is not extracted first, and leakage identification is performed directly on the entire image, it is easily affected by background thermal noise, lighting changes, and vegetation occlusion, thereby reducing the recognition accuracy.

[0124] In this embodiment, the infrared image corresponding to the main area of ​​the dam is input into a deep learning model. The deep learning model identifies candidate seepage areas and their seepage confidence levels. Specifically, the deep learning model can be a multi-scale feature extraction network integrating an attention enhancement mechanism, which can learn the local low-temperature anomalies, diffusion patterns, and boundary change features corresponding to seepage points from the infrared image. After the model runs, it outputs multiple candidate seepage areas and assigns a seepage confidence level to each candidate area to characterize the probability that the candidate area belongs to a real seepage point. It is understood that seepage in flood control dams usually manifests in infrared images as areas with low local temperatures, blurred edges, and irregular shapes. Relying solely on human experience is often insufficient for accurate judgment, while deep learning models can learn these complex features from a large number of labeled samples, thereby improving the initial candidate area extraction capability.

[0125] In this embodiment, the temperature difference features of the leakage candidate region in an infrared image are obtained. Specifically, for each leakage candidate region, a first average temperature of the pixels within the region is first calculated, then a preset range is expanded outside the region to form a background neighborhood, and a second average temperature of the background neighborhood is calculated. The difference between the two is then used as the temperature difference feature. Assuming the average temperature of the candidate region is... The average temperature of the background neighborhood is The temperature difference characteristic can then be expressed as: Understandably, the actual leakage area usually exhibits a more significant temperature difference than the surrounding area due to the increased water content. Especially under the combined effects of sunlight, evaporation, and wind speed, the temperature difference between the leakage area and the background area is more easily amplified. Therefore, temperature difference characteristics can provide important physical evidence for leakage determination.

[0126] In this embodiment, valid leak points are determined based on the temperature difference characteristics of the candidate leak areas, their geometric morphological features in visible light images, and the leak confidence level. Specifically, based on the location of the candidate leak areas in the visible light images, geometric morphological features are extracted, including the area of ​​connected regions, aspect ratio, and edge contour integrity. Subsequently, the temperature difference features, geometric morphological features, and leak confidence level are weighted and calculated to obtain a comprehensive evaluation score. When the comprehensive evaluation score is greater than a preset threshold, the candidate leak area is determined to be a valid leak point. It is understood that true leak points usually not only have significant thermal anomalies in infrared images but also exhibit a certain degree of morphological consistency in visible light images, such as extending along a slope, having a relatively continuous contour, or coinciding with the location of weak drainage areas. Therefore, by jointly determining the leak points based on temperature difference, geometric morphology, and confidence level, the accuracy of leak point screening can be effectively improved, and false alarms caused by local thermal noise can be reduced.

[0127] In this embodiment, environmental parameters during infrared image acquisition can be obtained, and an environmental correction coefficient can be determined based on these parameters to dynamically correct the leakage confidence level. Specifically, different correction functions are assigned to solar radiation intensity, wind speed, and ambient temperature. After obtaining the environmental correction coefficient, it is multiplied by the leakage confidence level to obtain the corrected leakage score. Then, the leakage candidate area is re-evaluated based on the corrected leakage score to suppress false thermal anomalies. It is understood that dam thermal maps are not only affected by leakage but also significantly interfered with by the external environment. For example, strong sunlight can raise the surface temperature, high wind speeds can accelerate surface heat dissipation, and changes in ambient temperature can cause the overall thermal field to drift. Therefore, introducing an environmental correction coefficient can make the judgment results more consistent with real inspection scenarios.

[0128] In this embodiment, leakage identification can also be performed on continuously acquired multi-frame infrared images to obtain the confidence sequence of the same leakage candidate region at different time points, and the confidence sequence is smoothed to calculate the time-series cumulative confidence. Specifically, if a candidate region persists in multiple frames of images, and its confidence sequence shows an overall increasing trend or remains stable, the final leakage confidence of the candidate region can be improved. It is understood that real leakage is usually persistent and evolving, and will not only occur sporadically in a single frame of image. False thermal anomalies caused by reflections, shadows, or local transient thermal disturbances are often short-lived and unstable. Therefore, combining multi-frame time-series information can further enhance the ability to identify real leakage points.

[0129] In this embodiment, the location information of the effective leakage point is output based on the registration relationship between the visible light image and the infrared image. Specifically, the pixel coordinates of the effective leakage point in the infrared image can be mapped to the corresponding position in the visible light image, and further combined with the geographical coordinates of the dam, inspection mileage, or electronic map information to output the location information of the effective leakage point in the actual dam scene. It is understood that infrared images are good at reflecting thermal anomalies, while visible light images are more suitable for describing actual spatial locations. Therefore, after coordinate mapping through the registration relationship, inspection personnel can quickly locate the leakage point on site and carry out verification, photography, marking, and repair.

[0130] In this embodiment, effective seepage points can be classified into risk levels based on their temperature difference intensity, area range, and spatially sensitive location within the main dam area. Corresponding graded early warning signals are then output based on the risk level, and these signals are associated with location information and stored in a disease database. Specifically, if an effective seepage point is located at the dam toe, joint area, or weak drainage area, a higher risk weight can be assigned; if the temperature difference is significant and the area expands rapidly, the early warning level can be further increased. It is understood that different seepage points have varying degrees of impact on dam safety. Combining the identification results with risk classification, early warning output, and database retention enables this method to not only have identification capabilities but also engineering management and operation and maintenance decision support capabilities.

[0131] This application provides a deep learning-based method for identifying seepage points in flood control dikes. This method integrates visible light images, infrared images, environmental parameters, and multi-frame temporal information to automatically identify, comprehensively determine, and accurately locate candidate seepage areas. Compared to traditional methods relying on manual inspections or single threshold judgments, this solution more effectively suppresses environmental interference, improves the ability to detect early, subtle seepage, reduces the probability of false alarms and missed alarms, and enables risk classification, early warning output, and defect archiving of seepage points. This enhances the intelligence level and engineering practical value of flood control dike inspections.

[0132] Figure 2 This is a schematic diagram of a module of a deep learning-based flood control dike seepage point identification system provided in an embodiment of this application. Figure 2 The deep learning-based flood control dike seepage point identification system 10 shown includes at least the following components: information acquisition module 11, feature extraction module 12, and result output module 13.

[0133] In this embodiment, the information acquisition module 11 is used to acquire visible light and infrared images of the dam area to be detected. Please refer to the following for details. Figure 1 The details and their corresponding descriptions are not repeated here.

[0134] In this embodiment, the feature extraction module 12 is used to extract the main body region of the dam based on the visible light image, input the infrared image corresponding to the main body region of the dam into a deep learning model, and identify leakage candidate regions and leakage confidence levels of the leakage candidate regions through the deep learning model; and obtain the temperature difference features of the leakage candidate regions in the infrared image. For details, please refer to [the relevant documentation / reference]. Figure 1 The details and their corresponding descriptions are not repeated here.

[0135] In this embodiment, the result output module 13 is used to determine the effective leakage point based on the temperature difference characteristics of the leakage candidate area, the geometric morphological characteristics in the visible light image, and the leakage confidence level; and to output the location information of the effective leakage point based on the registration relationship between the visible light image and the infrared image. Please refer to the following for details. Figure 1 The details and their corresponding descriptions are not repeated here.

[0136] Figure 3 This is an electronic device 20 provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device 20 includes at least the following components: a processor 21 and a memory 22.

[0137] In this embodiment, the memory 22 is used to store executable instructions of the processor 21, which, when configured to execute instructions, implement... Figure 1 The method shown in the figure.

[0138] In one embodiment of this application, the program operating in the electronic device 20 may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). The information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0139] It should be noted that a portion of the electronic device 20 described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0140] It should be noted that the term "computer" as used here refers to a computer built into electronic device 20, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into the computer.

[0141] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0142] It is understood that the deep learning-based flood control dike seepage point identification method, system and equipment provided in this application fuse visible light image and infrared image information. First, it accurately extracts the main area of ​​the dike based on the visible light image. Then, it combines the deep learning model to identify candidates for infrared thermal anomalies and introduces temperature difference features and geometric morphology features for secondary judgment, forming a two-stage discrimination mechanism of "model recognition + rule verification". This significantly improves the ability to identify early weak seepage anomalies and reduces false alarms and missed alarms caused by environmental interference.

[0143] Meanwhile, this application establishes a registration relationship between infrared and visible light images to achieve precise location of seepage points in actual dam scenarios. Combined with environmental correction coefficients and time-series confidence accumulation mechanisms, it effectively suppresses false thermal anomalies. On this basis, a risk level classification and early warning linkage mechanism is introduced so that the identification results can directly serve dam inspection and disease management, thereby improving the intelligence and engineering practicality of flood control dam safety monitoring.

[0144] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A deep learning-based method for identifying seepage points of flood control dams, characterized in that, The method includes: Acquire visible light and infrared images of the dam area to be inspected; Extract the main area of ​​the dam based on the visible light image; The infrared image corresponding to the main area of ​​the dam is input into a deep learning model, and the deep learning model is used to identify seepage candidate areas and the seepage confidence of the seepage candidate areas. Obtain the temperature difference characteristics of the candidate leakage area in the infrared image; Valid leakage points are determined based on the temperature difference characteristics of the candidate leakage areas, their geometric morphological characteristics in visible light images, and the leakage confidence level. Based on the registration relationship between the visible light image and the infrared image, the location information of the effective leakage point is output.

2. The method of claim 1, wherein, The step of determining the effective leakage point based on the temperature difference characteristics of the candidate leakage region, the geometric morphological characteristics in the visible light image, and the leakage confidence level includes: Based on the location of the leakage candidate region in the visible light image, the geometric morphological features are extracted, including the area of ​​the connected region, the aspect ratio, and the integrity of the edge contour. The temperature difference characteristics, the geometric morphology characteristics, and the leakage confidence level are weighted and calculated to obtain a comprehensive evaluation score. If the comprehensive evaluation score is greater than the preset score threshold, the candidate leakage area is determined to be a valid leakage point.

3. The method of claim 2, wherein, The step of obtaining the temperature difference characteristics of the leakage candidate region in the infrared image includes: Extract the first average temperature of the candidate leakage region in the infrared image; The background neighborhood of the candidate leakage area is determined after expanding outward by a preset range in the infrared image, and the second average temperature of the background neighborhood is extracted; The difference between the second average temperature and the first average temperature is used as the temperature difference characteristic of the leakage candidate area.

4. The method of claim 3, wherein, The deep learning model integrates an attention enhancement mechanism; The process of identifying leakage candidate regions using the deep learning model includes: The spatial attention module in the deep learning model is applied to enhance the weights of pixels with significant temperature differences in the main region of the dam. The channel attention module is used to extract thermal anomaly features at different scales from the infrared image to identify leakage morphology features.

5. The method of claim 4, wherein, The method further includes: Based on the temperature difference characteristic intensity, area range, and spatially sensitive location of the effective seepage point in the main area of ​​the dam, the effective seepage point is classified into risk levels. Based on the risk level, a corresponding graded early warning signal is output, and the graded early warning signal is associated with the location information and stored in the disease database of the dam area to be detected.

6. The method of claim 1, wherein, The step of determining the effective leakage point based on the temperature difference characteristics of the candidate leakage region, the geometric morphological characteristics in the visible light image, and the leakage confidence level includes: The environmental parameters during the infrared image acquisition are obtained, including sunlight intensity, wind speed, and ambient temperature. An environmental correction factor is determined based on the environmental parameters, and the leakage confidence level is dynamically corrected using the environmental correction factor.

7. The method of claim 6, wherein, The application of the environmental correction coefficient to dynamically correct the leakage confidence level includes: Multiply the leakage confidence level by the environmental correction factor to obtain the corrected leakage score; The candidate leakage regions are reassessed based on the corrected leakage score to suppress false thermal anomalies.

8. The method of claim 1, wherein, The method further includes: Leakage identification is performed on multiple consecutively acquired infrared images to obtain the confidence sequence of the same leakage candidate region at different time points; The confidence sequence is smoothed to calculate the time-series cumulative confidence of the leakage candidate region; If the time-series cumulative confidence of the leakage candidate region shows an increasing trend or remains stable, then the confidence of the leakage candidate region is increased.

9. A deep learning based leakage point identification system for flood embankment, applied to implement the method of any one of claims 1 to 8. The system includes: The information acquisition module is used to acquire visible light and infrared images of the dam area to be detected; The feature extraction module is used to extract the main body area of ​​the dam based on the visible light image, input the infrared image corresponding to the main body area of ​​the dam into a deep learning model, identify the leakage candidate area and the leakage confidence of the leakage candidate area through the deep learning model, and obtain the temperature difference feature of the leakage candidate area in the infrared image. The result output module is used to determine the effective leakage point based on the temperature difference characteristics of the leakage candidate area, the geometric morphology characteristics in the visible light image, and the leakage confidence level; and to output the location information of the effective leakage point based on the registration relationship between the visible light image and the infrared image.

10. An electronic device, comprising: include: processor; as well as A memory having computer-readable instructions stored thereon for controlling the processor to perform the method as described in any one of claims 1 to 8.

Images