All-weather water area hidden danger target monitoring method and device based on image enhancement

Abstract

The invention belongs to the technical field of vision and intelligent monitoring, and particularly relates to an all-weather water area hidden danger target monitoring method and device based on image enhancement. The method comprises the following steps: acquiring an all-weather water area scene image, marking a coordinate frame of a hidden danger target by using a marking tool, and generating an all-weather water area hidden danger target data set; correcting the acquired all-weather water area scene image by using an extreme weather correction algorithm; adopting a yo11-x model as a hidden danger target identification model of the water area scene; and deploying an extreme weather correction algorithm and the trained yolo11-x model to a server side, and if a hidden danger target is recognized, pushing target information to an application side. The problem of imaging degradation of fog, rain and night scenes is solved, and the problem that the real-time requirement is difficult to meet due to too high calculation load is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of visual and intelligent monitoring, and more specifically, relates to a method and device for monitoring potential water hazards based on image enhancement. Background Technology

[0002] Waterway safety monitoring is a crucial aspect of water conservancy project management, port operations, and public safety protection. Traditional waterway monitoring primarily relies on manual patrols or fixed video surveillance systems. However, these methods face significant challenges in practical applications. Weather conditions severely restrict monitoring effectiveness: atmospheric scattering in foggy conditions leads to blurry images; raindrops and motion blur in rainy weather affect target identification; and conventional cameras struggle to acquire effective information in low-light conditions at night, while infrared illumination is easily interfered with by water surface reflections. Manual monitoring suffers from inefficiency, with personnel prone to fatigue leading to missed detections, and it is difficult to achieve full coverage and real-time response across large water areas.

[0003] Chinese patent document CN106296612A discloses a graded surveillance video dehazing system and method guided by image quality evaluation and weather conditions, including a video acquisition module, a video conversion module, a video image processing module, and a display module. The video acquisition and conversion modules acquire images and convert them into digital image frames. The video image processing module integrates weather conditions and image quality evaluation to determine the image dehazing level, and applies corresponding algorithms to dehaze images of different levels.

[0004] Existing image processing-based technologies also have limitations. Traditional motion detection methods, such as Gaussian mixture models, are sensitive to dynamic water wave interference and are prone to high false alarm rates; mainstream dehazing algorithms, such as dark channel priors, are prone to halo artifacts in dynamic water scenes; low-light enhancement techniques, such as histogram equalization, amplify noise and lose details while increasing brightness. Although deep learning solutions have made some progress in recent years, they still face key bottlenecks: most models are only effective under single ideal weather conditions, and their performance drops sharply in complex environments such as fog, rain, and night; multi-sensor fusion solutions, such as visible light + infrared, improve robustness, but the computational load is too high to meet real-time requirements. Summary of the Invention

[0005] This invention aims to overcome at least one of the shortcomings of the prior art and proposes an all-weather adaptive monitoring method for potential water hazards based on image enhancement. By constructing an image enhancement pipeline that integrates a physical model and deep learning, the imaging degradation problem in foggy, rainy, and nighttime scenes is effectively solved; and the lightweight target detection algorithm YOLOv11 is used to improve the water target recognition capability.

[0006] The detailed technical solution of this invention is as follows: A method for monitoring potential water hazards based on image enhancement, the method comprising: S1. Acquire all-weather water scene images and use annotation tools to annotate the coordinate boxes of potential hazards, generating an all-weather water hazard target dataset for training the hazard target recognition model; S2. Use extreme weather correction algorithms to correct the collected all-weather water scene images, including extreme weather judgment and image enhancement for various extreme weather conditions, and construct a test set of water scene images; S3. The YOLO11-X model is used as the target identification model for water scene. It is trained using an all-weather water scene target dataset. After training, the target identification is performed on the water scene image set to be tested. S4. Deploy the extreme weather correction algorithm and the trained YOLO11-X model to the server. If a potential hazard is identified, push the target information to the application.

[0007] Furthermore, the extreme weather assessment specifically includes: S211. Use an atmospheric scattering physical model to determine whether the acquired water scene image is foggy: First, based on the atmospheric scattering model, the image with fog is determined to be: I(x)=J(x)t(x)+A(1 t(x))(1) In formula (1), I(x) is the observed foggy image, J(x) is the radiation of the fog-free scene, A is the global atmospheric light, and t(x) is the transmittance. The fog concentration scoring function is constructed as follows: (2); In formula (2), ω(x) is a local window centered at x, Ω is the entire image region, x is the input image, and y is the pixel in the local window of ω(x). The atmospheric scattering model value calculated for a single channel c; Then, a fog concentration score is used to determine a threshold for whether it is a foggy day: (3) In formula (3), To determine the scene type, As an empirical threshold, The standard deviation of the image gradient. The contrast threshold is used; Foggy represents foggy weather, and Normal represents normal weather. S212. Use a joint frequency-spatial analysis method to determine whether the acquired water scene image is from a rainy day: First, calculate the energy characteristics of the rain streaks: (4); In formula (4), Let be the rain pattern feature value, i,j be the coordinates of a point in the image, M,N be the image width and height, and I() be the grayscale value calculated. x and y is the Sobel operator: (5); Then, the color dispersion feature is calculated, namely the standardized standard deviation of the histograms of the Hue and Saturation channels in the HSV space: (6); In formula (6), The calculated color dispersion characteristic value, and These are histograms for the Hue and Saturation channels, respectively, with std() being the normalization operation. Finally, use the following formula to determine whether it is a rainy day: (7); In formula (7), For dynamic thresholds, Here, k represents the mean of the color dispersion, and k is the steepness coefficient of the Sigmoid curve. The characteristic value of rain ripple energy. This represents the theoretical estimate or empirical upper limit of the maximum rain ripple gradient energy that the algorithm is expected to detect in the current scenario. S213. Use a brightness statistical analysis-based method to determine whether the acquired water scene image is nighttime: The characteristic of the proportion of dark areas can be expressed as: (8) In formula (8), Igray(i,j) is the pixel value of the grayscale image at coordinates (i,j), and τ dark =50 is the threshold for dark pixels. H , W These are the image's length and width, respectively. N Total number of pixels; Brightness distribution skewness for: (9); In formula (9), μ bright The average brightness of the grayscale image. σ bright The standard deviation of a grayscale image is σ , μ 3 represents the third-order center distance; The rules for determining nighttime are based on brightness, proportion of dark areas, bias, and saturation in the HSV color space. Represented as: (10); In formula (10), β 1 represents the upper limit threshold for overall brightness. β 2 is the lower threshold for the proportion of dark areas. β 3 is the lower limit threshold for skewness. It is the saturation channel in the HSV color space. S The average value, β 4 is the upper limit threshold for the mean saturation value; S214. Process the judgment image using a time-series method: First, a time window with a duration of 5 seconds is established. When the indicator of more than 70% of the frames in the time window is judged to be extreme weather, the result of that weather is output. Conversely, if more than 70% of the frames in the time window do not meet the requirements for extreme weather, the current image enhancement state will be exited.

[0008] Furthermore, the image enhancement for various extreme weather conditions specifically includes: S221, Foggy Image Enhancement; The LD-Net algorithm is used to dehaze the image, which includes eight convolutional layers and three fully connected layers. The atmospheric scattering physical model is transformed based on LD-Net, converting the two parameters t and A into a single parameter K. The value of K is predicted using a neural network. After obtaining K, the enhanced fog-free image J(x) is calculated using formula (11): (11); (12); In formula (11)-(12), This is the deviation value, which defaults to 1.

[0009] Furthermore, depthwise separable convolutions are used to replace the ordinary convolutional layers in LD-Net.

[0010] S222, Rainy Day Image Enhancement: The PReNet algorithm is used to remove rain from the image. PReNet first uses a shallow residual network of 5 ResBlocks as the basis, and then expands a ResNet repeatedly in T stages. The network parameters are reused in different stages. Then, a recurrent layer is added on top of this. The loop takes the output of stage t (fin) and the state of the previous recurrent layer as input, and is implemented using LSTM. The final expression is as follows: (13); (14); (15); In formula (13)-(15), For Conv+ReLU, receive the image output from stage t-1. and the original rain map The concatenation of features is used as input, and the feature maps are fused to obtain intermediate results. ; and These are the cyclic features of the current state and the previous state, respectively. For 5 ResBlocks, extract deep feature representations; This is a convolutional layer, and its output is the rain removal result. This is the output of the current stage t, and the output of the last stage is the final enhanced rainy day image.

[0011] Furthermore, to improve the efficiency of the rain removal network, a lightweight modification was made to PReNet, replacing the ordinary convolutional layers in ResBlock with depthwise separable convolutions, which reduced the number of parameters and computational cost of the rain removal model and improved its efficiency; at the same time, ReLU in ResBlock was replaced with the SiLU activation function.

[0012] S223. Nighttime Image Enhancement: The SCI algorithm is used to enhance the low-light performance of the image. According to Retinex theory, a low-light observed image is equal to the dot product of a clear image and the illumination, i.e.: (16); In formula (16), n is the clear image, m is the light intensity of the image, and p is the low-light observation image; Then, a progressive lighting optimization process was constructed using the SCI algorithm: (17); In formula (17), To optimize the network, n t+1 Let u be the illumination intensity at time t+1. t Let n be the illumination residual at stage t. t For the illumination during phase t, H θ This is an illumination estimation network, built from a convolutional neural network, n 0Let P be the initial value of illumination; SCI includes multi-stage output and incorporates a self-calibration module. : (18); In formula (18), For division operation, For a series of convolutional neural networks, v is the output of the convolutional neural network at the current stage. t It is calibrated and used for the next stage. The input, and finally the low-light enhancement network becomes ; Through the final low-light enhancement network The optimization yields the final light intensity m. * Low-light image p divided by m * The final enhanced nighttime image n is obtained. * .

[0013] Furthermore, in order to improve the operating efficiency of the low-light enhancement network, SCI will be used... and Part of the convolutional neural network was replaced with depthwise separable convolutions, resulting in an efficient low-light enhancement network.

[0014] Furthermore, the YOLOv11-x model will be improved: Insert an LSKA module into the backbone network of yolo11-x, that is, replace the PSA attention structure in the C2PSA module with an LSKA module; LSKA employs a depthwise separable convolution design.

[0015] In another aspect of the present invention, an image enhancement-based all-weather water hazard monitoring device is provided, the device comprising: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform an image enhancement-based all-weather water hazard target monitoring method as described above.

[0016] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform the image enhancement-based all-weather water hazard target monitoring method described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention provides an all-weather water area hidden danger target monitoring method and device based on image enhancement. It adopts an all-weather image correction and target recognition method based on deep learning, which can automatically complete the monitoring of hidden danger targets in water areas without human intervention. It is more efficient than manual inspection and is suitable for real-time monitoring of large-scale water areas. It can realize intelligent monitoring of water areas in all weather conditions by combining with video monitoring equipment. There is no need to install additional complex sensors or equipment. It has low cost and good scalability.

[0018] (2) The present invention provides an all-weather water area hidden danger target monitoring method and device based on image enhancement. Through adaptive image defogging, deraining and low light enhancement algorithms, it can perform image enhancement and target recognition for water area scenes under various extreme weather conditions, and has strong environmental adaptability. At the same time, it can quickly detect hidden danger targets in water area scenes and trigger alarms through similarity thresholds, providing timely and effective decision support for the safety supervision and maintenance of water areas. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an image enhancement-based all-weather water hazard monitoring method according to the present invention.

[0020] Figure 2 This is a comparison image before and after processing by the dehazing algorithm in Embodiment 1 of the present invention.

[0021] Figure 3 This is a comparison image before and after the rain removal algorithm processing in Embodiment 1 of the present invention.

[0022] Figure 4 This is a comparison image before and after low-light enhancement algorithm processing in Embodiment 1 of the present invention.

[0023] Figure 5 This is an example diagram of the water scene recognition effect in Embodiment 1 of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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 invention pertains.

[0026] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0028] Embodiment 1 Refer Figure 1 , this embodiment provides an all-weather water area hidden danger target monitoring method based on image enhancement, and the method includes: S1: Construction of an all-weather water area hidden danger target data set: Obtain all-weather water area scene images and use annotation tools to annotate the coordinate frames of hidden danger targets, and generate an all-weather water area hidden danger target data set for training the hidden danger target recognition model.

[0029] Preferably, the actual operation scene images containing hidden danger targets such as fishing, fishing, and sand mining are collected by a camera installed on the water area monitoring device. The collected images include various extreme weather conditions in addition to sunny days, such as rainy days, foggy days, and nights. Then, use the Labelimg annotation tool to annotate the coordinate frames of hidden danger targets from the collected images, including the positions and categories of hidden danger targets. Finally, generate an all-weather water area hidden danger target data set containing various water area hidden danger targets. There are a total of 3000 annotated images, including 2400 in the training set and 600 in the validation set. The all-weather water area hidden danger target data set is used for training the hidden danger target recognition model in S3.

[0030] S2: Water area image enhancement in extreme weather, that is, an extreme weather correction algorithm: Use the extreme weather correction algorithm to correct the collected all-weather water area scene images, including extreme weather judgment and image enhancement for various extreme weather conditions, and construct a water area scene image test set.

[0031] Preferably, the extreme weather judgment specifically includes judging whether the all-weather water area scene image is a foggy day, a rainy day, a dark night, or a sunny day. If it is a foggy day, a rainy day, or a dark night, the following processing is performed; if it is a sunny day, directly perform hidden danger target recognition: S211. For the water area scene image obtained from the monitoring device, use the atmospheric scattering physical model to judge whether the image is a foggy day, and if it is a foggy day, perform image defogging.

[0032] First, according to the atmospheric scattering model, a foggy image can be expressed as: I(x)=J(x)t(x)+A(1 t(x))(1) In formula (1), I(x) is the observed foggy image, J(x) is the radiation of the fog-free scene, A is the global atmospheric light, and t(x) is the transmittance. The fog concentration scoring function is: (2); In formula (2), ω(x) is a local window centered at x, typically 15×15 in size, Ω is the entire image region, x is the input image, and y is the pixel in the local window of ω(x). The atmospheric scattering model value calculated for a single channel c; Use fog concentration scoring to determine a threshold for identifying whether it is a foggy day: (3); In formula (3), To determine the scene type, This is an empirical threshold, with a typical value of 0.65; The standard deviation of the image gradient; The contrast threshold is 25, with a typical value of 25; Foggy represents foggy weather, and Normal represents normal weather. When fog concentration is scored and If the current image is foggy, it is determined to be foggy; otherwise, it is determined to be normal weather.

[0033] S212. Use a joint frequency-spatial analysis method to determine whether the acquired water scene image is from a rainy day; if it is from a foggy or rainy day, perform image deraining: First, calculate the energy characteristics of the rain streaks: (4); In formula (4), Let be the feature value of the rain pattern, i,j be the coordinates of a point in the image, M,N be the width and height of the image, and I() be the calculation of the grayscale image. x, y is the Sobel operator: (5); Then, the color dispersion feature is calculated, namely the standardized standard deviation of the histograms of the Hue and Saturation channels in the HSV space: (6); In formula (6), The calculated color dispersion characteristic value, and These are histograms for the Hue and Saturation channels, respectively, with std() being the normalization operation. Finally, use the following formula to determine whether it is a rainy day: (7); In formula (7), This is a dynamic threshold, typically 0.6; The mean of color dispersion; k is the steepness coefficient of the Sigmoid curve; The characteristic value of rain ripple energy; This represents the theoretical estimate or empirical upper limit of the maximum rain ripple gradient energy that the algorithm is expected to detect in the current scenario.

[0034] S213. For the acquired water scene image, use a brightness statistical analysis method to determine whether the image is in darkness. If it is in darkness, perform low-light enhancement on the image. The characteristic of the proportion of dark areas can be expressed as: : (8); In formula (8), I gray ( i , j ) is a grayscale image in coordinates ( i , j The pixel value at position τ dark =50 (8-bit image) is the dark pixel threshold. H , W These are the image's length and width, respectively. N Total number of pixels; Brightness distribution skewness for: (9); In formula (9), μ bright The average brightness of the grayscale image. σ bright The standard deviation of a grayscale image is σ , μ 3 represents the third-order center distance; The rules for determining nighttime are based on brightness, proportion of dark areas, bias, and saturation in the HSV color space. Represented as: (10); In formula (10), β 1 is the upper limit threshold for overall brightness, set to 80; β 2 is the lower limit threshold for the proportion of dark areas, which is set to 0.7; β 3 is the lower threshold of skewness, which is set to 1.3; It is the saturation channel in the HSV color space. S The average value; β 4 is the upper limit threshold for the mean saturation value, which is set to 0.3.

[0035] S214. In order to reduce misjudgments in extreme weather conditions such as fog, rain, and darkness, a time-series method is used to process the judgment image.

[0036] First, establish a time window with a duration of 5 seconds. When more than 70% of the frames in the time window are judged to be extreme weather, output the result for that weather.

[0037] Conversely, if more than 70% of the frames in the time window do not meet the requirements for extreme weather, the current image enhancement state will be exited.

[0038] If the weather is determined to be foggy, execute S221; if the weather is determined to be rainy, execute S222; if the weather is determined to be nighttime, execute S223.

[0039] Preferably, image enhancement is performed for various extreme weather conditions, specifically including: S221. Foggy Image Enhancement: When an image is determined to be foggy, the foggy image enhancement algorithm is activated to remove the fog from the image.

[0040] This embodiment uses the Light-DehazeNet (LD-Net) algorithm to dehaze the image. LD-Net is a lightweight dehazing algorithm based on deep learning, which includes eight convolutional layers and three fully connected layers.

[0041] The atmospheric scattering physics model is transformed using LD-Net, converting the two-parameter t and A into a single-parameter K, and then a neural network is used to predict the value of K. ,get The enhanced haze-free image was then calculated using formula (11). : (11); (12); In formula (11)-(12), This is the bias value, with a default value of 1. The LD-Net method can effectively dehaze water images. In practical applications, it is used in conjunction with subsequent recognition algorithms. To improve the running efficiency of LD-Net, depthwise separable convolutions are used to replace the ordinary convolutional layers in LD-Net.

[0042] like Figure 2As shown, the images are comparisons of the water area foggy day images before and after defogging enhancement provided by the present invention. The left image is the water area foggy day image without defogging enhancement, and the right image is the water area foggy day image after defogging enhancement. It can be seen that the defogging enhancement effect is very obvious.

[0043] S222, Rainy Day Image Enhancement: When an image is determined to be rainy, the rainy day image enhancement algorithm is activated to remove the rain from the image.

[0044] This embodiment uses the PReNet algorithm to remove rain from images. PReNet first uses a shallow residual network of 5 ResBlocks as a base, and then expands a ResNet repeatedly over a total of T stages, with the network parameters being reused in different stages.

[0045] The inference process at each stage is described by the following formula: (19); (20); Then, a recurrent layer, Frecurrent, was added on top of this. The recurrent layer takes the output of the current stage's fin and the state of the previous stage's recurrent layer as input, and is implemented using LSTM. The recurrent layer can propagate the relationships between features from different stages. The final expression becomes as follows: (13); (14); (15); In formula (13)-(15), For Conv+ReLU, receive the image output from stage t-1. and the original rain map The concatenation of features is used as input, and the feature maps are fused to obtain intermediate results. ; and These are the cyclic features of the current state and the previous state, respectively. For 5 ResBlocks, extract deep feature representations; This is a convolutional layer, and its output is the rain removal result. This is the output of the current stage t, and the output of the last stage is the final enhanced rainy day image.

[0046] To improve the efficiency of the rain removal network, a lightweight modification was made to PReNet, replacing the ordinary convolutional layers in ResBlock with depthwise separable convolutions, which reduced the number of parameters and computational cost of the rain removal model and improved its efficiency. In addition, to retain more effective information, the ReLU activation function in ResBlock was replaced with the SiLU activation function.

[0047] like Figure 3 As shown, this is a comparison image of a water area image before and after rain enhancement provided by an example of the present invention. The left image is the unenhanced rain image, and the right image is the enhanced rain image, with significantly improved clarity.

[0048] S223, Nighttime Image Enhancement: When an image is determined to be nighttime, the low-light nighttime image enhancement algorithm is activated to enhance the low-light performance of the image.

[0049] This embodiment uses the SCI algorithm to enhance the low-light performance of the image. According to Retinex theory, a low-light observed image is equal to the dot product of a sharp image and the illumination, that is: (16); In formula (16), n is the clear image, m is the light intensity of the image, and p is the low-light observation image; Therefore, illumination estimation is usually considered the primary optimization objective. Once accurate illumination is obtained, a clear image can be directly derived from the aforementioned relationship. Inspired by the staged illumination optimization process in existing work, the SCI algorithm constructs a progressive illumination optimization process: (17); In formula (17), To optimize the network, n t+1 Let u be the illumination intensity at time t+1. t Let n be the illumination residual at stage t. t For the illumination during phase t, H θ This is an illumination estimation network, built from a convolutional neural network, n 0 Let P be the initial value of illumination.

[0050] SCI includes multi-stage outputs. To ensure that the outputs at different stages of the training process converge to the same state, a self-calibration module is added. : (18); In formula (18), For division operation, A series of convolutional neural networks, i.e., the number of convolutional neural networks is defined according to actual needs. v is the output of the convolutional neural network at the current stage. t It is calibrated and used for the next stage. The input, and finally the low-light enhancement network becomes ; Through the final low-light enhancement network The optimization yields the final light intensity m. *Low-light image p divided by m * The final enhanced nighttime image n is obtained. * To improve the operating efficiency of the low-light enhancement network, Hθ and... Part of the convolutional neural network was replaced with depthwise separable convolutions, resulting in an efficient low-light enhancement network.

[0051] SCI uses lightweight neural networks to achieve efficient low-light image enhancement, such as... Figure 4 The image shown is a comparison of the low-light enhancement algorithm provided in this invention before and after processing. The left image is the unenhanced night image, and the right image is the enhanced night image, with significantly improved clarity.

[0052] S3. Hazardous target identification in enhanced water area images: The YOLO11-X model is used as the hazardous target identification model for water area scenes. It is trained using an all-weather water area hazardous target dataset. After training, hazardous targets are identified on the test set of water area scene images.

[0053] The specific improvements in YOLOv11 are as follows: LSKA (Large Kernel Separable Convolutional Attention) is inserted into the backbone network of YOLOv11-x. In water monitoring scenarios, there are many small and occluded targets. The attention mechanism can acquire global image information, effectively improving the detection accuracy of small and occluded targets. Therefore, adding LSKA has a better effect on identifying potential hazards in water scenes. In addition, LSKA uses a depthwise separable convolutional design, which requires less computation and does not affect the recognition efficiency of the target detection network.

[0054] First, the YOLO11-X model is trained. In the actual application of the algorithm, the image is first input into step S2 for image processing, and then sent to the YOLO11-X model for target recognition. The input image is the enhanced image obtained in step S2, with an input resolution of 1280×1280. The main targets to be identified are potential hazards such as anglers, fishermen, illegal swimmers, sand dredgers, excavators, and floating foreign objects in water scenes.

[0055] like Figure 5 This example provides a representation of the target identification results for potential hazards in aquatic scenes, as provided in this invention. The text above the detection box indicates the identification category and confidence level: algae, fishing rod, angler, fishing net, and bucket. The target identification results for potential hazards in aquatic scenes detected by this invention are consistent with the actual annotation results.

[0056] S4. Water Hazard Target Alarm Service: Deploy the extreme weather correction algorithm and the trained YOLO11-X model to the server. If a hazard target is identified, the target information will be pushed to the application.

[0057] The extreme weather correction algorithm in S2 and the YOLO11-x model trained in S3 are deployed to the server, and an inference service is set up on the server port. Under normal circumstances, the inference service is in a listening state. After receiving images and equipment information transmitted from the monitoring device, it first performs extreme weather judgment and image enhancement on the image, and then sends the preprocessed image to the target detection model. The detection result is obtained through inference by the YOLO11-x model. If the result contains a potential target, the target information is pushed to the application.

[0058] Example 2 This embodiment provides an apparatus for implementing an image enhancement-based all-weather monitoring method for potential water hazards, the apparatus comprising: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform an image enhancement-based all-weather water hazard target monitoring method as described above.

[0059] In this embodiment, electronic devices include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile computing devices, smartphones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld devices, messaging devices, wearable computing devices, consumer electronic devices, etc.

[0060] Example 3 This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed, cause the machine to perform the image enhancement-based all-weather water hazard target monitoring method described above.

[0061] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0062] In this case, the program code itself, which can be read from the readable medium, can perform the functions of any of the above embodiments, and therefore the computer-readable code and the readable storage medium storing the computer-readable code constitute a part of this specification.

[0063] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for monitoring potential water hazards based on image enhancement, characterized in that, The method includes: S1. Acquire all-weather water scene images and use annotation tools to annotate the coordinate boxes of potential hazards, generating an all-weather water hazard target dataset for training the hazard target recognition model; S2. Use extreme weather correction algorithms to correct the collected all-weather water scene images, including extreme weather judgment and image enhancement for various extreme weather conditions, and construct a test set of water scene images; S3. The YOLO11-X model is used as the target identification model for water scene. It is trained using an all-weather water scene target dataset, and then target identification is performed on the water scene image test set. S4. Deploy the extreme weather correction algorithm and the trained YOLO11-X model to the server. If a potential hazard is identified, push the target information to the application.

2. The image enhancement-based all-weather water hazard monitoring method according to claim 1, characterized in that, The extreme weather assessment specifically includes: S211. Use an atmospheric scattering physical model to determine whether the acquired water scene image is foggy: First, based on the atmospheric scattering model, the image with fog is determined to be: I(x)=J(x)t(x)+A(1 t(x))(1); In formula (1), I(x) is the observed foggy image, J(x) is the radiation of the fog-free scene, A is the global atmospheric light, and t(x) is the transmittance. The fog concentration scoring function is constructed as follows: （2）； In formula (2), ω(x) is a local window centered at x, Ω is the entire image region, x is the input image, and y is the pixel in the local window of ω(x). The atmospheric scattering model value calculated for a single channel c; Then, a fog concentration score is used to determine a threshold for whether it is a foggy day: （3）； In formula (3), To determine the scene type, As an empirical threshold, The standard deviation of the image gradient. The contrast threshold is used; Foggy represents foggy weather, and Normal represents normal weather. S212. Use a joint frequency-spatial analysis method to determine whether the acquired water scene image is from a rainy day: First, calculate the energy characteristics of the rain streaks: （4）； In formula (4), Let be the rain pattern feature value, i,j be the coordinates of a point in the image, M,N be the image width and height, and I() be the grayscale value calculated. x and y is the Sobel operator: （5）； Then, the color dispersion feature is calculated, namely the standardized standard deviation of the histograms of the Hue and Saturation channels in the HSV space: （6）； In formula (6), The calculated color dispersion characteristic value, and These are histograms for the Hue and Saturation channels, respectively, with std() being the normalization operation. Finally, use the following formula to determine whether it is a rainy day: （7）； In formula (7), For dynamic thresholds, Here, k represents the mean of the color dispersion, and k is the steepness coefficient of the Sigmoid curve. The characteristic value of rain ripple energy. This represents the theoretical estimate or empirical upper limit of the maximum rain ripple gradient energy that the algorithm is expected to detect in the current scenario. S213. Use a brightness statistical analysis-based method to determine whether the acquired water scene image is nighttime: The characteristic of the proportion of dark areas can be expressed as: （8）； In formula (8), Igray(i,j) is the pixel value of the grayscale image at coordinates (i,j), and τ dark =50 is the threshold for dark pixels. H , W These are the image's length and width, respectively. N Total number of pixels; Brightness distribution skewness for: （9）； In formula (9), μ bright The average brightness of the grayscale image. σ bright The standard deviation of a grayscale image μ 3 represents the third-order center distance; The rules for determining nighttime are based on brightness, proportion of dark areas, bias, and saturation in the HSV color space. Represented as: （10）； In formula (10), β 1 represents the upper limit threshold for overall brightness. β 2 is the lower threshold for the proportion of dark areas. β 3 is the lower limit threshold for skewness. It is the saturation channel in the HSV color space. S The average value, β 4 is the upper limit threshold for the mean saturation value; S214. Process the judgment image using a time-series method: First, a time window with a duration of 5 seconds is established. When the indicator of more than 70% of the frames in the time window is judged to be extreme weather, the result of that weather is output. Conversely, if more than 70% of the frames in the time window do not meet the requirements for extreme weather, the current image enhancement state will be exited.

3. The image enhancement-based all-weather water hazard monitoring method according to claim 2, characterized in that, The image enhancement for various extreme weather events specifically includes: S221. Foggy Image Enhancement: The LD-Net algorithm is used to dehaze the image, including eight convolutional layers and three fully connected layers. The atmospheric scattering physics model is transformed using LD-Net, converting the two-parameter t and A into a single-parameter K, and then a neural network is used to predict the value of K. ,get The enhanced haze-free image was then calculated using formula (11). : （11）； （12）； In formula (11)-(12), This is the deviation value; S222, Rainy Day Image Enhancement: The PReNet algorithm is used to remove rain from the image. PReNet first uses a shallow residual network of 5 ResBlocks as the basis, and then expands a ResNet repeatedly in T stages. The network parameters are reused in different stages. Then, a recurrent layer is added on top of this. The loop takes the output of stage t (fin) and the state of the previous recurrent layer as input, and is implemented using LSTM. The final expression is as follows: （13）； （14）； （15）； In formula (13)-(15), For Conv+ReLU, receive the image output from stage t-1. and the original rain map The concatenation of features is used as input, and the feature maps are fused to obtain intermediate results. ; and These are the cyclic features of the current state and the previous state, respectively. For 5 ResBlocks, extract deep feature representations; This is a convolutional layer, and the output is the rain removal result. This is the output of the current stage t; the output of the last stage is the final enhanced rainy day image. S223. Nighttime Image Enhancement: The SCI algorithm is used to enhance the low-light performance of the image. According to Retinex theory, a low-light observed image is equal to the dot product of a clear image and the illumination, i.e.: （16）； In formula (16), n is the clear image, m is the light intensity of the image, and p is the low-light observation image; Then, a progressive lighting optimization process was constructed using the SCI algorithm: （17）； In formula (17), To optimize the network, n t+1 Let u be the illumination intensity at time t+1. t Let n be the illumination residual at stage t. t For the illumination during phase t, H θ This is an illumination estimation network, built from a convolutional neural network, n 0 Let P be the initial value of illumination; SCI includes multi-stage output and incorporates a self-calibration module. : （18）； In formula (18), For division operation, For a series of convolutional neural networks, v is the output of the convolutional neural network at the current stage. t It is calibrated and used for the next stage. The input, and finally the low-light enhancement network becomes Through the final low-light enhancement network The optimization yields the final light intensity and the final enhanced nighttime image.

4. The image enhancement-based all-weather water hazard monitoring method according to claim 1, characterized in that, Replace the regular convolutional layers in LD-Net with depthwise separable convolutions.

5. The image enhancement-based all-weather water hazard monitoring method according to claim 3, characterized in that, Based on PReNet, depthwise separable convolutions are used to replace the ordinary convolutional layers in ResBlock; at the same time, ReLU in ResBlock is replaced with the SiLU activation function.

6. The all-weather water hazard monitoring method based on image enhancement according to claim 3, characterized in that, SCI and Some convolutional neural networks were replaced with depthwise separable convolutions.

7. The image enhancement-based all-weather water hazard monitoring method according to claim 3, characterized in that, The YOLOv11-x model will be improved as follows: An LSKA module is inserted into the backbone network of YOLOv11-x. LSKA uses a depthwise separable convolutional design.

8. An apparatus for implementing an image enhancement-based all-weather monitoring method for potential water hazards, characterized in that, The device includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements the steps of the image enhancement-based all-weather water hazard target monitoring method as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images