Water surface oil stain detection method, system and equipment

Abstract

The invention belongs to the technical field of oil stain detection, and particularly relates to a water surface oil stain detection method, system and equipment, the water surface oil stain detection system comprises a flexible circuit board FPC, an InGaAs infrared sensor, an infrared narrow-band filter and an optical lens, light enters a visible light camera and the optical lens through an external lens, the visible light camera collects visible light images, and the visible light images are transmitted to the InGaAs infrared sensor; the optical lens focuses a light beam, then light with a specific wavelength is screened out through the near-infrared narrow-band filter, the screened light is received by the InGaAs infrared sensor, and the InGaAs infrared sensor is connected with the FPC; the method has the advantages that visible light imaging and short wave infrared spectrum analysis technologies are ingeniously fused, so that the accuracy and the reliability of oil stain detection are remarkably improved. Under the current background that environmental protection and water resource management are increasingly severe, powerful technical support is undoubtedly provided for water surface pollution monitoring.

Description

Technical Field

[0001] This application belongs to the technical field of oil pollution detection, and particularly relates to a method, system and device for detecting oil pollution on the water surface. Background Art

[0002] With the rapid development of society, the problem of aging industrial facilities has become increasingly prominent, especially metal structure pipelines such as oil pipelines, sewage pipelines and natural gas transportation pipelines. These pipelines are exposed to water vapor and corrosive substances in the transported raw materials for a long time, and inevitably corrosion and damage will occur, resulting in leakage accidents, and then triggering serious ecological and environmental crises.

[0003] In response to this problem, there are mainly the following solutions: One is to use new materials and new technologies to improve the durability of pipelines. However, due to the large number of already buried pipelines, replacing them all along the line is not only costly but also may cause waste of resources. The second is to conduct real-time monitoring by burying sensors. However, the effectiveness of this method highly depends on the accuracy of the sensors. Usually, only when the pipeline is severely damaged resulting in pollution or environmental changes, the sensors will trigger an alarm, showing obvious hysteresis. Especially in the application of underwater pipelines, once a leakage occurs, if it cannot be detected in time, it will bring huge environmental risks.

[0004] Therefore, it is urgent to develop more efficient and accurate monitoring technologies and combine preventive maintenance strategies to reduce the environmental and economic risks brought by pipeline aging and leakage. Summary of the Invention

[0005] Based on the above problems, this application uses multi-spectral imaging technology, combines visible light and short-wave infrared spectra, and realizes efficient detection of oil pollution on the water surface. Its technical solution is as follows: A method for detecting oil pollution on the water surface includes the following steps: S1. Data acquisition and preprocessing; S2. Construct the graph structure of the oil film area; S3. Construct a dual-link detection model, including a main detection network and a graph-guided network. Use the main detection network to detect the visible light image in real time, and model the spectral-spatial correlation characteristics of the oil film through the graph-guided network to realize oil film pollution monitoring.

[0006] Preferably, for data acquisition in step S1: Since the oil film shows significantly different physical properties in the visible light and infrared bands, it is necessary to ensure the spatio-temporal alignment of the dual-modal data. The method is as follows: S101. Arrange heatable metal dots on the calibration plate; S102. Synchronously collect visible light images and infrared thermal images; S103. Realize pixel-level alignment by solving the perspective transformation matrix: ; Among them, is the visible light image pixel coordinate, is the corresponding pixel coordinate of the infrared image, and H is a 3×3 homography matrix (solved by the least squares method); Preferably, the data acquisition preprocessing is as follows: Differentiated preprocessing is adopted according to the characteristics of different modality data: Visible light channel: The improved MSRCR algorithm is used to enhance the iridescence effect of the oil film: ; Among them, is the pixel value of the enhanced visible light image at the coordinate (x, y), is the weight coefficient (default is 0.33), is the pixel value of the original visible light image at the coordinate (x, y), is the Gaussian kernel, is the enhancement intensity parameter (generally 0.5), is the color restoration factor; Infrared band channel: Oil film characteristic peak: ; is the difference window, is the infrared wave wavelength.

[0007] Preferably, the steps of constructing the graph structure of the oil film area in step S2 are as follows: S21. Node generation strategy: S210. Initial segmentation: Input the visible light image into the lightweight DeepLabv3+ network to obtain the pixel-level segmentation mask, and at the same time use the infrared spectral angle mapping algorithm to calculate the similarity between each pixel and the standard oil film spectrum: ; Among them, is the infrared spectral vector of pixel p, is the standard oil film spectrum template vector; represents the norm of the vector; Take the area where the SAM value < 0.2 rad as the candidate oil film area; S211. Superpixel clustering: Apply the improved SLIC superpixel algorithm to the candidate area, introducing the spectral distance constraint: ; Among them, is the Euclidean distance in the color space, is the pixel coordinate distance, is the Euclidean distance of the infrared band vector, and m, s, and n are normalization coefficients.

[0008] S22. Edge connection mechanism: S220: The establishment of edges needs to consider both spatial adjacency and spectral similarity: Spatial constraint: Use Delaunay triangulation to establish the initial topology and set the maximum edge length; Spectral verification: For each candidate edge , calculate its infrared feature correlation coefficient: ; Among them, is the infrared feature vector of node i, is the infrared feature vector of node j, represents the covariance of the infrared feature vectors of nodes i and j and ; and represent the standard deviations of the infrared feature vectors of nodes i and j respectively; Retain the edges of to suppress false connections; S221: Dynamic update: Re-evaluate the graph structure every M frames and adopt a node merging strategy to reduce the computational complexity: = ; Among them, and are the spatial coordinates of nodes and node respectively, represents the Euclidean distance of the position coordinates of nodes and .

[0009] Preferably, in step S3, the main detection network adopts the YOLOv8 architecture, adds a spectral attention sublayer to its RGB branch, and designs a spectral pyramid pooling on the IR branch to convert the 128-dimensional spectral vector into multi-scale features. The expression is as follows: ; Among them, represents the connection operation, which splices the features after different operations; is the maximum pooling operation on the spectral vector v, and the pooling kernel size is 1; is the average pooling operation on the spectral vector v, and the pooling kernel size is 3; is the convolution operation on the spectral vector v, and the convolution kernel size is 5.

[0010] Introduce a cross-modal attention mechanism in the FPN layer: ; Among them, represents the activation function, which maps the input value to the interval (0, 1). is the visible light feature map, is the infrared feature map, represents the convolution operation, which is used to process the concatenated features.

[0011] The detection head outputs key points for oil film contour point prediction.

[0012] Preferably, in step S3, the graph guiding network adopts the GMM model, and the specific method is as follows: Node encoding layer: Use a 3-layer GATv2 network, with the number of attention heads per layer , and the node update formula: ; Among them, represents the updated node 's feature vector; is the activation function, which performs a non-linear transformation on the summation result; represents the set of neighbor nodes of node ; is the attention coefficient, calculated through the gating mechanism; is the trainable weight matrix, used to transform the feature vectors of neighbor nodes; is the feature vector of node .

[0013] Graph pooling layer: Adopt the TopKPooling method to retain the 30% nodes with the highest importance: ; is the importance score of node , used to select and retain nodes with high importance during graph pooling; is the trainable vector, used to calculate the importance of nodes; is the feature vector of node ; represents the norm of vector ; Graph readout layer: Generate a graph-level representation through the dynamic routing algorithm: ; Graph structure sparsification: Adopt the K-NN strategy to limit the maximum number of edges of each node; Among them, is the total number of nodes in the graph; represents the importance coefficient of node , which is obtained by calculation and is used to measure the contribution degree of node to the graph-level representation; represents the eigenvector of node ; is a trainable vector used to calculate the importance coefficient of the node; is the transpose of vector ; represents the exponential operation.

[0014] Preferably, it further includes S4 early warning, and a three-level early warning mechanism is established: Monitoring level Level1: Record events when ThreatScore > 0.5; Warning level Level2: Trigger an audible and visual alarm when Score > 0.7 and lasts for more than 3 frames; Emergency level Level3: Automatically start the pollution source tracing program when Score > 0.9.

[0015] An oil pollution detection and monitoring system for water surfaces includes a data acquisition module, a data processing module, and a data output module; Data acquisition module: Collect visible light and infrared spectral data and calibrate the data; Data processing module: Preprocess the collected three-dimensional data, construct the graph structure of the oil film area, and perform oil film pollution analysis using a double-link detection model; Data output module: Give early warnings about the oil pollution detection and monitoring results.

[0016] An oil pollution detection and monitoring device for water surfaces includes an InGaAs infrared sensor, a near-infrared narrowband filter, and an optical lens. Light enters the visible light camera and the optical lens respectively through an external lens. The visible light camera collects visible light images, and the optical lens focuses the light beam. Subsequently, light of a specific wavelength is filtered out by the near-infrared narrowband filter, and the filtered light is received by the InGaAs infrared sensor. The InGaAs infrared sensor is connected to a flexible circuit board FPC.

[0017] Preferably, the visible light images collected by the camera can provide the graph and boundary information of the oil pollution, while the short-wave infrared spectrum can further identify the type and concentration of the oil pollution through the differences in spectral reflection and absorption intensities; the visible light camera and the optical lens need to zoom synchronously to ensure the synchronism of the imaging.

[0018] Compared with the prior art, the beneficial effects of this application are as follows: 1. Water and oil films have different reflection characteristics for light of different wavelengths. By analyzing the variation of reflection intensity with wavelength, it can be determined that the difference in reflection intensity between water and oil films is most significant within a specific wavelength range. Selecting this wavelength band for filtering can effectively distinguish the reflected light of water and oil films, thereby achieving precise detection of oil films.

[0019] 2. The synchronous zoom mechanism realizes the synchronous zoom of two groups of lenses through gear connection, ensuring the synchronism and consistency of imaging. This design not only improves the detection accuracy but also reduces misjudgment caused by inconsistent focal lengths.

[0020] 3. The high-precision data analysis technology can automatically process visible light and infrared spectral data through multi-spectral data fusion and algorithm recognition, and comprehensively judge by combining the visible light characteristics and infrared characteristics of oil stains. This high-precision data analysis technology significantly improves the accuracy and reliability of detection. Brief Description of the Drawings

[0021] Figure 1 It is a schematic structural diagram of this application.

[0022] 1 - Optical lens, 2 - Near-infrared narrow-band filter, 3 - InGaAs infrared sensor, 4 - Flexible printed circuit board FPC, 5 - Lens, 6 - Lens mounting base. Detailed Description of the Embodiment

[0023] The technical solution of this application will be described in detail below through specific embodiments and the accompanying drawings. It should be understood that the specific features in the embodiments of this application are a detailed description of the technical solution of this application, rather than a limitation on the technical solution of this application. The specific technical features can be combined with each other.

[0024] An equipment for detecting and monitoring oil stains on the water surface includes a visible light camera and a near-infrared camera. The near-infrared camera includes an InGaAs infrared sensor 3, a near-infrared narrow-band filter 2, and an optical lens 1. Light enters the visible light camera and the optical lens 1 respectively through an external lens. The visible light camera captures visible light imaging. The light beam focused by the optical lens 1 then passes through the near-infrared narrow-band filter 2 to screen out light of a specific wavelength. The filtered light enters the lens 5 and is received by the InGaAs infrared sensor 3. The InGaAs infrared sensor 2 is connected to the flexible printed circuit board FPC.

[0025] The lens 5 is mounted on the lens mounting base 6, and the near-infrared narrow-band filter 2 can be attached to the lens 5.

[0026] The visible light imaging captured by the visible light camera can provide the graph and boundary information of the oil spill, while the near-infrared camera can further identify the type and concentration of the oil spill through the differences in spectral reflection and absorption intensity; the visible light camera and the optical lens need to be zoomed synchronously, ensuring the synchronicity and consistency of the imaging.

[0027] A water surface oil spill detection and monitoring system includes a data acquisition module, a data processing module, and a data output module; Data acquisition module: Collect visible light and infrared spectrum data and calibrate the data; Data processing module: Preprocess the collected three-dimensional data, construct the graph structure of the oil film area, and perform oil film pollution analysis using the double-link detection model; Data output module: Give early warnings about the results of the oil spill detection and monitoring.

[0028] By capturing the spectral differences between the oil film and the water body in the infrared band, the near-infrared camera can more accurately identify the type and composition of the oil spill. Different types of oil spills will exhibit different absorption and reflection characteristics in the infrared band, and these characteristics can be captured and analyzed by the near-infrared camera, thus realizing the accurate identification of the oil spill.

[0029] When the visible light camera detects an abnormal area on the water surface, the near-infrared camera will perform a zoom operation synchronously with the visible light camera to conduct multi-spectral analysis on the suspected leakage point. Since the oil film and the water body have significantly different spectral reflection and absorption characteristics in the short-wave infrared band, the near-infrared camera can accurately identify the spectral characteristics of the oil spill by capturing these differences. For example, crude oil will exhibit obvious absorption peaks in specific infrared bands, while the water body may exhibit a higher reflectivity. By comparing and analyzing the data in the visible light and short-wave infrared bands, the system can more comprehensively identify the presence and characteristics of the oil spill.

[0030] As attached Figure 1As shown in the figure, the core structure and working principle of the near-infrared camera are detailed. When light enters the camera from the external environment, it first passes through the external lens and then is focused by the optical lens system to ensure that the light can accurately converge on the photosensitive area of the sensor. The focused light then passes through the near-infrared narrow-band filter, which is used to screen out near-infrared light of specific wavelengths and filter out stray light in other bands, thus ensuring that only light of the target wavelength can enter the subsequent sensor part. This filtering process is crucial for improving the imaging accuracy and signal-to-noise ratio. The filtered near-infrared light is then received by the InGaAs infrared sensor. The InGaAs sensor is a high-performance infrared detector with advantages such as high sensitivity, wide spectral response range, and low noise, and is particularly suitable for detection in the near-infrared band. After receiving the optical signal, the sensor converts it into an electrical signal and connects it to the FPC (flexible printed circuit board) through gold wires. As a high-conductivity and low-resistance connection material, gold wires can ensure the stability and reliability of the signal during transmission. The other end of the FPC is connected to a computer or other data processing device, which is responsible for transmitting the data collected by the sensor to the network or local system for further processing. This design not only simplifies the structure of the camera but also improves the flexibility and scalability of the system. In this way, the near-infrared camera can efficiently capture, transmit, and process near-infrared light signals and is widely used in fields such as night vision, security monitoring, medical imaging, and industrial inspection. The structural design of this near-infrared camera fully considers the coordinated work of optics, electronics, and mechanics, ensuring its high-performance performance in complex environments. Through the organic combination of the optical lens, narrow-band filter, InGaAs sensor, and FPC, the camera can efficiently complete the acquisition, transmission, and processing of optical signals, providing a reliable solution for various application scenarios.

[0031] In practical applications, the coordinated working mechanism of this system demonstrates powerful detection capabilities. The system can monitor the water surface status in real time through the visible light camera. Once an abnormal area is detected, the near-infrared camera immediately conducts multi-spectral analysis to determine the type and scope of the oil pollution. This coordinated working mechanism not only improves the detection efficiency but also significantly reduces the false alarm rate, especially under complex environmental conditions such as light changes and water surface fluctuations. The system can also, through the coordinated work of the visible light camera (visible light camera) and the short-wave infrared camera (near-infrared camera), monitor the water surface status in real time and promptly detect and handle oil spill incidents. This real-time monitoring and rapid response mechanism can not only effectively prevent the spread of oil pollution but also minimize environmental pollution and economic losses. The coordinated working mechanism of the visible light camera and the short-wave infrared camera can greatly improve the accuracy and reliability of detection.

[0032] All the data collected by the visible light camera and the short-wave infrared camera will be aggregated and input into an efficient data analysis algorithm. These algorithms combine the characteristics of oil on the water surface in the visible light and infrared bands, and through machine learning or deep learning models, conduct multi-dimensional analysis on the data. For example, the algorithm can extract the color and texture features of the oil film in the visible light image, and at the same time combine features such as spectral reflectance and absorptance in the short-wave infrared image to construct a comprehensive discrimination model. This multi-modal data analysis method can not only improve the accuracy of oil pollution detection, but also effectively reduce the false alarm rate, especially in complex environmental conditions (such as light changes, water surface fluctuations, etc.).

[0033] In addition, the collaborative working mechanism of the system is also reflected in the real-time processing and feedback of data. When suspected oil pollution is detected, the system will immediately trigger an alarm and upload the relevant data to the monitoring center for further analysis and decision-making. At the same time, the system can also automatically adjust the camera parameters (such as focal length, exposure time, etc.) according to the detection results to optimize the acquisition quality of subsequent data. This intelligent detection process not only improves the detection efficiency, but also provides reliable data support for subsequent pollution control. Through the collaborative work of the visible light camera and the short-wave infrared camera, combined with advanced data analysis algorithms, the system can achieve high-precision and real-time detection of oil on the water surface. This multi-spectral fusion detection method not only overcomes the limitations of single-band detection, but also provides an efficient and reliable solution for water surface oil pollution monitoring, and has broad application prospects in fields such as environmental protection and industrial safety.

[0034] A method for detecting oil on the water surface includes the following steps: S1. Data acquisition and preprocessing; In the data acquisition of step S1: Since the oil film exhibits significantly different physical properties in the visible light and infrared bands, it is necessary to ensure the spatio-temporal alignment of the dual-modal data. The method is as follows: S101. Arrange heatable metal dots on the calibration plate; S102. Synchronously acquire visible light images and infrared thermal images; S103. Realize pixel-level alignment by solving the perspective transformation matrix: ; Wherein, is the original pixel intensity value, represents, is the weight coefficient (default is 0.33), is the pixel value of the original visible light image at the coordinate (x, y), is the Gaussian kernel, is the enhancement intensity parameter (generally 0.5), is the color restoration factor; The data acquisition preprocessing is as follows: Differentiated preprocessing is adopted according to the characteristics of different modal data: Visible light channel: The improved MSRCR algorithm is used to enhance the iridescence effect of the oil film: ; Among them, is the pixel value at the coordinate (x, y) of the enhanced visible light image, is the weight coefficient (default is 0.33), is the Gaussian kernel, is the enhancement intensity parameter (usually 0.5), is the color restoration factor; Infrared band channel: Oil film characteristic peak: ; is the differential window, is the infrared wave wavelength.

[0035] S2. Construct the graph structure of the oil film area; The steps of constructing the graph structure of the oil film area in step S2 are as follows: S21. Node generation strategy: S210. Initial segmentation: Input the visible light image into the lightweight DeepLabv3+ network to obtain the pixel-level segmentation mask, and at the same time use the infrared spectral angle mapping algorithm to calculate the similarity between each pixel and the standard oil film spectrum: ; Among them, is the infrared spectral vector of pixel p, is the standard oil film spectrum template vector; represents the norm (modulus length) of the vector.

[0036] Take the area where the SAM value < 0.2rad as the candidate oil film area; S211. Superpixel clustering: Apply the improved SLIC superpixel algorithm to the candidate area, introducing spectral distance constraints: ; Among them, is the Euclidean distance in the color space, is the pixel coordinate distance, is the Euclidean distance of the infrared band vector, and m, s, n are normalization coefficients.

[0037] S22. Edge connection mechanism: S220: The establishment of edges needs to consider both spatial adjacency and spectral similarity: Spatial constraint: Use Delaunay triangulation to establish the initial topology and set the maximum side length; Spectral verification: For each candidate edge , calculate its infrared feature correlation coefficient: ; where, is the infrared feature vector of node i, is the infrared feature vector of node j, represents the covariance of the infrared feature vectors of nodes i and j and ; and represent the standard deviations of the infrared feature vectors of nodes i and j respectively.

[0038] Retain edges to suppress false connections; S221: Dynamic update: Re-evaluate the graph structure every M frames and adopt a node merging strategy to reduce the computational complexity: = .

[0039] where, and are the spatial coordinates of nodes and node respectively, represents the Euclidean distance between the position coordinates of nodes and .

[0040] S3. Construct a dual-link detection model, including a main detection network and a graph-guided network. Use the main detection network to perform real-time detection on visible light images, and model the spectral-spatial correlation characteristics of the oil film through the graph-guided network to achieve oil film pollution monitoring.

[0041] In step S3, the main detection network adopts the YOLOv8 architecture, adds a spectral attention sublayer to its RGB branch, and designs a spectral pyramid pooling on the IR branch to convert the 128-dimensional spectral vector into multi-scale features. The expression is as follows: ; where, represents the concatenation operation, which concatenates the features after different operations; is the max pooling operation on the spectral vector v, and the pooling kernel size is 1; is the average pooling operation on the spectral vector v, and the pooling kernel size is 3; is the convolution operation on the spectral vector v, and the convolution kernel size is 5.

[0042] Introduce a cross-modal attention mechanism in the FPN layer: ; Among them, represents the activation function, which maps the input value to the interval (0, 1). is the visible light feature map, is the infrared feature map, represents the convolution operation, which is used to process the concatenated features.

[0043] The detection head outputs key points for oil film contour point prediction.

[0044] In step S3, the graph-guided network adopts the GMM model, and the specific method is as follows: Node encoding layer: Use a 3-layer GATv2 network, with the number of attention heads in each layer , and the node update formula: ; Among them, represents the updated node 's feature vector; is the activation function, which performs a non-linear transformation on the sum result; represents the set of neighbor nodes of node ; is the attention coefficient, calculated through the gating mechanism; is the trainable weight matrix, which is used to transform the feature vectors of neighbor nodes; is the feature vector of node .

[0045] Graph pooling layer: Adopt the TopKPooling method to retain the 30% nodes with the highest importance: ; is the importance score of node , which is used to select and retain nodes with high importance during graph pooling; is the trainable vector, which is used to calculate the importance of nodes; is the feature vector of node ; represents the norm (modulus length) of vector .

[0046] Graph readout layer: Generate a graph-level representation through the dynamic routing algorithm: ; Graph structure sparsification: The k-NN strategy is adopted to limit the maximum number of edges of each node; Among them, is the total number of nodes in the graph; represents the importance coefficient of node , which is obtained by calculation and is used to measure the contribution degree of node to the graph-level representation; represents the eigenvector of node ; is a trainable vector used to calculate the importance coefficient of the node; is the transpose of vector ; represents the natural constant, which is used for exponential operation here.

[0047] Among them, is the total number of nodes in the graph; represents the importance coefficient of node i, which is obtained by calculation and is used to measure the contribution degree of node i to the graph-level representation; represents the eigenvector of node i; is a trainable vector used to calculate the importance coefficient of the node; is the transpose of vector u; represents the natural constant, which is used for exponential operation here.

[0048] S4 Early warning, establish a three-level early warning mechanism: Monitoring level Level1: Record events when ThreatScore > 0.5; Early warning level Level2: Trigger audible and visual alarms when Score > 0.7 and lasts for more than 3 frames; Emergency level Level3: Automatically start the pollution source tracing program when Score > 0.9.

[0049] To sum up, the water surface oil pollution detection system described in this patent application realizes high-precision and real-time detection of water surface oil pollution through the collaborative work of visible light cameras and short-wave infrared cameras, combined with advanced data analysis algorithms. This system not only overcomes the limitations of single-band detection, but also provides an efficient and reliable solution for water surface oil pollution monitoring. In the fields of environmental protection, industrial safety, etc., this system has broad application prospects and huge market potential. With the continuous progress of technology and the continuous deepening of applications, it is believed that this system will play an increasingly important role in future water surface oil pollution monitoring and contribute more to protecting our water resources and environmental safety.

[0050] It should be noted that the design and implementation of this system involve cutting-edge knowledge in multiple technical fields, including optical imaging technology, image processing technology, machine learning algorithms, etc. During the R & D process of the system, we fully drew on the latest research results and practical experiences in related fields at home and abroad, and made innovative improvements and optimizations in combination with actual needs. Therefore, this system not only has high technological advancement and practicality, but also has good scalability and maintainability, and can meet the application requirements in different scenarios.

[0051] The above are only the preferred embodiments of this application. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the technical principle of this application, several improvements and deformations can still be made, and these improvements and deformations should also be regarded as the protection scope of this application.

Claims

1. A method for detecting oil pollution on the water surface, characterized in that, Including the following steps: S1. Data acquisition and preprocessing; S2. Construct the graph structure of the oil film area; S3. Construct a dual-link detection model, including a main detection network and a graph-guided network. Use the main detection network to detect visible light images in real time, and model the spectral-spatial correlation characteristics of the oil film through the graph-guided network to achieve oil film pollution monitoring.

2. The method for detecting oil pollution on the water surface according to claim 1, wherein, In step S1, data acquisition: Since the oil film exhibits significantly different physical characteristics in the visible and infrared bands, it is necessary to ensure the spatio-temporal alignment of the dual-modal data. The method is as follows: S101. Arrange heatable metal dots on the calibration plate; S102. Synchronously collect visible light images and infrared thermal images; S103. Achieve pixel-level alignment by solving the perspective transformation matrix: ； Among them, is the pixel coordinate of the visible light image, is the corresponding pixel coordinate of the infrared image, and H is a 3×3 homography matrix.

3. A method for detecting oil pollution on the water surface according to claim 1, characterized in that, The data acquisition preprocessing is as follows: Adopt differential preprocessing according to the characteristics of different modal data: Visible light channel: Use the improved MSRCR algorithm to enhance the iridescence effect of the oil film; ； wherein, is the pixel value of the enhanced visible light image at the coordinate (x, y), is the weight coefficient, is the pixel value of the original visible light image at the coordinate (x, y), is the Gaussian kernel, is the enhancement intensity parameter, is the color restoration factor; Infrared band channel: Oil film characteristic peak: ； is a differential window, is the wavelength of the infrared wave.

4. The water surface oil pollution detection method according to claim 1, characterized in that, The steps for constructing the graph structure of the oil film area in step S2 are as follows: S21. Node generation strategy: S210. Initial segmentation: Input the visible light image into the lightweight DeepLabv3+ network to obtain a pixel-level segmentation mask, and at the same time use the infrared spectral angle mapping algorithm to calculate the similarity between each pixel and the standard oil film spectrum: ； Among them, is the infrared spectral vector of pixel p, is the standard oil film spectral template vector; represents the norm of the vector; Take the area with SAM value <0.2rad as the candidate oil film area; S211. Superpixel clustering: Apply the improved SLIC superpixel algorithm to the candidate area, introducing spectral distance constraints: ； Among them, is the Euclidean distance of the color space, is the pixel coordinate distance, is the Euclidean distance of the infrared band vector, and m, s, and n are normalization coefficients; S22. Edge connection mechanism: S220: When establishing edges, both spatial adjacency and spectral similarity need to be considered: Spatial constraint: Use Delaunay triangulation to establish the initial topology and set the maximum side length; Spectral verification: For each candidate edge , calculate its infrared feature correlation coefficient: ； Among them, is the infrared feature vector of node i, is the infrared feature vector of node j, represents the covariance of the infrared feature vectors of nodes i and j and ; and represent the standard deviations of the infrared feature vectors of nodes i and j respectively; Keep the edges to suppress spurious connections; S221: Dynamic update: Re-evaluate the graph structure every M frames and adopt a node merging strategy to reduce the computational complexity: = ； Among them, and are the spatial coordinates of nodes and node respectively, represents the Euclidean distance of the position coordinates of nodes and respectively.

5. A method for detecting oil pollution on the water surface according to claim 1, characterized in that, In step S3, the main detection network adopts the YOLOv8 architecture, and a spectral attention sublayer is added to its RGB branch. On the IR branch, a spectral pyramid pooling is designed to convert the 128-dimensional spectral vector into multi-scale features. The expression is as follows: ； Among them, represents a concatenation operation, which concatenates the features after different operations; performs a max pooling operation on the spectral vector v; performs an average pooling operation on the spectral vector v; performs a convolution operation on the spectral vector v; Introduce a cross-modal attention mechanism in the FPN layer: ； Among them, represents an activation function that maps the input value to the interval (0, 1), is the visible light feature map, is the infrared feature map, represents a convolution operation used to process the concatenated features; The detection head outputs key points for predicting the oil film contour points.

6. The method for detecting oil pollution on the water surface according to claim 1, wherein, In step S3, the graph-guided network adopts the GMM model. The specific method is as follows: Node encoding layer: Using a 3-layer GATv2 network, the number of attention heads in each layer , node update formula: ； Among them, represents the feature vector of the updated node ; is the activation function, which performs a non-linear transformation on the summation result; represents the set of neighbor nodes of node ; is the attention coefficient, which is calculated through a gating mechanism; is a trainable weight matrix, which is used to transform the feature vectors of neighbor nodes; is the feature vector of node ; Graph pooling layer: Adopt the TopKPooling method to retain the 30% nodes with the highest importance: ； is the node importance score, used to select nodes with high importance when performing graph pooling; is a trainable vector used to calculate the importance of nodes; is the node feature vector; represents the norm of the vector ; Graph readout layer: Generate a graph-level representation through a dynamic routing algorithm: ； Graph structure sparsification: Adopt the K-NN strategy to limit the maximum number of edges of each node; Among them, is the total number of nodes in the graph; represents the node importance coefficient, represents the node eigenvector; is a trainable vector used to calculate the importance coefficient of the node; is the vector transpose; represents the exponential operation.

7. A method for detecting oil pollution on the water surface according to claim 6, characterized in that, It also includes S4 early warning, establishing a three-level early warning mechanism: Monitoring level Level1: Record events when ThreatScore>0.5; Warning level Level2: Trigger an audible and visual alarm when Score>0.7 and lasts for more than 3 frames; Emergency level Level3: Automatically start the pollution source tracing program when Score>0.

9.

8. An oil pollution detection and monitoring system for water surface, characterized in that, Adopt the water surface oil pollution detection and monitoring method described in any one of claims 1-7, including a data acquisition module, a data processing module, and a data output module; Data acquisition module: Collect visible light and infrared spectral data and calibrate the data; Data processing module: preprocess the collected three-dimensional data, construct the graph structure of the oil film area, and perform oil film pollution analysis using the dual-link detection model; Data output module: give early warnings about the results of oil pollution detection and monitoring.

9. An oil pollution detection and monitoring device for water surface, characterized in that, The water surface oil pollution detection and monitoring method according to any one of claims 1-7 is adopted, including an InGaAs infrared sensor, a near-infrared narrow-band filter, and an optical lens. Light enters the visible light camera and the optical lens respectively through the external lens. The visible light camera collects visible light imaging, and the optical lens focuses the light beam. Subsequently, light of a specific wavelength is filtered out by the near-infrared narrow-band filter, and the filtered light is received by the InGaAs infrared sensor. The InGaAs infrared sensor is connected to the flexible circuit board FPC.

10. The water surface oil pollution detection and monitoring device according to claim 9, characterized in that, The visible light imaging collected by the visible light camera can provide the graph and boundary information of the oil pollution, while the short-wave infrared spectrum can further identify the type and concentration of the oil pollution through the differences in spectral reflection and absorption intensities; the visible light camera and the optical lens need to zoom synchronously to ensure the synchronism of the imaging.

Images