Water surface spilled oil automatic recovery system, water surface spilled oil identification model and method, and medium

Abstract

The invention belongs to the field of water surface oil pollutant treatment, and particularly relates to a water surface spilled oil automatic recovery system, a water surface spilled oil identification model, a method and a medium, in the system, an unmanned cruise device cruises in a target water area and is positioned by an automatic positioning navigation subsystem; the image acquisition device acquires water surface images in real time during cruising and transmits the water surface images to the automatic control subsystem, and the automatic positioning navigation subsystem acquires corresponding position information of the images; the automatic control subsystem identifies an oil stain area in the water surface image through an integrated intelligent oil-water image identification module, and transmits the water surface image with the oil stain area to the automatic positioning navigation subsystem; the automatic positioning navigation subsystem positions the actual position of an oil contamination area in the image according to the position information corresponding to the water surface image, then automatically plans the optimal oil contamination recovery path of the unmanned cruise device, and guides the unmanned cruise device to go to a designated water area; and after arrival, the oil stain is automatically recovered through the oil recovery device. The robot is high in automation level and high in operation precision and efficiency in a complex water area environment.

Description

Technical Field

[0001] This invention belongs to the field of water surface oil pollutant treatment technology, specifically relating to an automatic oil spill recovery system and an oil spill identification model, method, and medium. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of the marine transportation industry, oil spills occur frequently. The spread of oil in the ocean can cause losses and harm to the ecological environment, economic development, marine industry, and the lives of coastal residents.

[0004] Traditional oil collection equipment can generally only collect oil spills within a fixed area, which greatly limits its application scope and flexibility. Furthermore, it relies heavily on manual observation and operation, which not only increases the complexity of operation but also significantly limits its efficiency when dealing with large or dynamically changing oil spill areas, making it difficult to meet the needs of efficient and rapid cleaning.

[0005] Therefore, some new technologies have emerged:

[0006] 1. Chinese patent document CN104908897B describes an autonomous patrolling oil spill recovery robot that uses an unmanned vessel as a carrier and is equipped with an advanced oil spill detection device to achieve automated oil spill recovery operations. However, this solution has relatively high detection costs for polluted areas; the lack of integrated intelligent algorithms results in insufficient detection accuracy; and the large size of the equipment not only increases the difficulty of transportation and deployment but also limits its application flexibility in narrow or specific scenarios.

[0007] 2. Chinese patent document CN107401149B describes a method for recovering oil spills on the water surface. The method involves manually operating unmanned boats and drones to locate and collect oil spills on the water surface. Long-term manual remote monitoring and operation can easily cause visual fatigue and distraction, which in turn affects the accuracy and efficiency of the operation. The manual operation of the collection operation is highly complex and it is difficult to respond quickly and make precise adjustments, resulting in a significant reduction in the efficiency of oil spill recovery.

[0008] 3. Chinese patent document CN1 13920445A describes a method for detecting oil spills on the sea surface based on multi-kernel classification model decision fusion. This method integrates satellite remote sensing technology and uses a multi-kernel classification model decision fusion method to achieve effective detection of oil spills on the sea surface. However, this scheme mainly focuses on the monitoring and identification of oil spills on the sea surface, and does not further integrate related functions for the actual cleanup and collection of oil pollution.

[0009] 4. Chinese patent document CN113177929B describes a method, system and application for detecting oil spills on the sea surface based on a fully convolutional network. The image acquisition device is mounted on a relatively traditional UAV, and a feature pyramid module is introduced. The mounting platform determines the angle of image capture and environmental conditions, which in turn affects the implementation details of the detection scheme, algorithm design and specific development of the technical route. Summary of the Invention

[0010] To overcome the shortcomings of the prior art, this invention provides an automatic oil spill recovery system, an oil spill identification model, method, and medium, which not only improves the automation level of the oil recovery equipment but also significantly enhances its operational accuracy and efficiency in complex aquatic environments, thereby better meeting the actual needs of water body environmental protection and crude oil recovery.

[0011] The technical solution adopted in this invention is: an automatic oil spill recovery system for water surfaces, the system comprising an unmanned surface patrol vehicle and an oil recovery device, an image acquisition device, an automatic positioning and navigation subsystem, and an automatic control subsystem installed on the unmanned surface patrol vehicle;

[0012] The image acquisition device acquires water surface images in real time and sends the water surface images to the automatic control subsystem;

[0013] The automatic control subsystem identifies the oil spill area based on the water surface image and instructs the automatic positioning and navigation subsystem to obtain the location information of the oil spill area;

[0014] The unmanned surface patrol vehicle arrives at the oil spill area based on the location information and uses the oil recovery device to perform the oil spill recovery task.

[0015] Furthermore,

[0016] The water surface image carries the location information captured by the unmanned surface patrol vehicle.

[0017] This invention also provides a water surface oil spill recognition model, applied to the aforementioned automatic water surface oil spill recovery system. The automatic control subsystem integrates the water surface oil spill recognition model and uses it to identify oil-contaminated areas based on the water surface image. The water surface oil spill recognition model includes:

[0018] A multi-layer feature extraction module is used to extract visual features from the input water surface image, capture oil slick features, and extract spatial information of significant areas. The visual features include local details and global background.

[0019] The graph convolutional neural network module is used to enrich the contextual information in the visual features of each salient region and to extract the contextual information between multiple salient regions.

[0020] The attention mechanism module is used to assign weights to the visual features based on their correlation with oil and water, highlighting the most indicative areas in the water surface image, so that the oil spill recognition model can focus on the most relevant visual features for prediction, thereby identifying the oil spill in the water surface image and its location.

[0021] Furthermore,

[0022] The multi-layer feature extraction module includes:

[0023] The feature extraction network, based on a deep convolutional neural network, is used to extract visual features and capture oil slick features from the input water surface image and extract feature maps.

[0024] The RPN module is used to dynamically generate multiple anchor points on the feature map, predict the probability score of the anchor point belonging to the oil pollution on the water surface and the position offset relative to the anchor point for each anchor point, and then propose a set of high-probability candidate regions based on the predicted probability score and position offset. Each candidate region is a suggestion, and a preliminary binary classification is performed on each suggestion to determine whether there is oil pollution and its approximate location. The candidate regions with oil pollution are designated as salient regions.

[0025] The RoI pooling module is used to extract RoI features from the feature map, adjust these RoI features of different sizes, convert them into feature sub-maps of uniform and preset size, and map the feature sub-maps back to the entire feature map.

[0026] The R-CNN module is used to output the entire feature map obtained by the RoI pooling module to the fully connected layer of the R-CNN module itself to perform oil pollution prediction and location regression. The oil pollution prediction is to identify whether there is oil pollution in the water surface image, and the location regression is to predict and locate the position of the oil pollution in the water surface image.

[0027] Furthermore,

[0028] A global attention mechanism is introduced into the feature extraction network. The global attention mechanism includes: building bridges between feature maps of adjacent resolutions and extracting global contextual information from low-resolution feature maps to enhance the weights of high-resolution features.

[0029] Furthermore,

[0030] The feature extraction network employs a 13-layer convolutional neural network and utilizes 3×3 convolutional kernels. The feature extraction network consists of five core convolutional modules, each of which sequentially outputs feature maps of different dimensions.

[0031] Furthermore,

[0032] The RPN module and the R-CNN module share the convolutional layers of the entire feature map.

[0033] Furthermore,

[0034] The oil spill detection model is trained using a multi-task loss function, which comprehensively considers both classification and localization tasks. The formula for this multi-task loss function is as follows:

[0035]

[0036] in,

[0037] This represents the probability that the i-th anchor point predicted by the oil spill identification model is the target.

[0038] r * i This represents the true category label of the i-th anchor point, which is 1 when a target exists and 0 otherwise.

[0039] S i and S i * Let represent the regression predicted value and the actual value of the position of the i-th anchor point, respectively;

[0040] N cls and N reg These are the normalized values ​​for the classification and regression tasks, used to balance the loss contributions of the two tasks;

[0041] L cls and L reg These are the classification loss and regression loss functions, respectively.

[0042] p i * It is a binary variable that is 1 when the i-th anchor point is a positive example and 0 otherwise. It is used to control that only positive examples participate in the calculation of regression loss.

[0043] λ is a hyperparameter used to adjust the weighting between classification loss and regression loss.

[0044] The present invention also provides a training method for an oil spill detection model, used to train the aforementioned oil spill detection model, the training method comprising:

[0045] Acquire online images of oil spills on the water surface;

[0046] An image dataset is composed of oil-water samples selected from the online images, the oil-water samples including various oil-water types;

[0047] The image dataset is divided into a training sample set and a test sample set, wherein the ratio of the training sample set to the test sample set is set to 4:1.

[0048] The oil-water samples in the training sample set are labeled;

[0049] The labeled oil-water samples are input into a pre-built oil spill identification model for training.

[0050] The optimized oil spill recognition model was obtained after training.

[0051] The present invention also provides a method for identifying oil spills on water surfaces, based on the aforementioned oil spill identification model, the method comprising:

[0052] Acquire images of the target water surface area as images to be identified;

[0053] The image to be identified is input into the trained oil spill recognition model, and the recognition information of the image to be identified is output. The recognition information includes classification information and location information. The classification information is whether there is oil in the image to be identified, and the location information is the specific location of the oil in the image to be identified.

[0054] The present invention also provides a computer-readable storage medium storing one or more programs, characterized in that, when the one or more programs are executed, they can implement the aforementioned oil spill identification model training method or the aforementioned oil spill identification method.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. It has improved the automation level of oil recovery equipment and significantly enhanced its operational accuracy and efficiency in complex aquatic environments, thereby better meeting the actual needs of water environment protection and crude oil recovery;

[0057] 2. By selecting a new platform, the unmanned surface patrol vehicle, to carry the image acquisition device, a different observation perspective than previous platforms is provided, and the system is given the ability to directly participate in surface operations. In view of the different perspectives and environmental conditions brought about by this, this invention combines the unmanned surface patrol vehicle with the surface oil spill identification model, which not only realizes the all-round monitoring of surface oil pollution, but also provides a practical solution for subsequent oil pollution cleanup, demonstrating its great potential in the field of marine environmental protection.

[0058] 3. The oil spill identification model for water surfaces employs a multi-layer feature extraction module and a graph convolutional neural network, and embeds an attention mechanism module into the model, providing a comprehensive and accurate method for detecting oil spills on water surfaces. It can maintain high precision and accuracy. The algorithm of this oil spill identification model can accurately analyze images and identify oily areas on the water surface, providing target guidance for subsequent oil recovery operations.

[0059] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings.

[0060] The invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of an automatic oil spill recovery system according to an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the structure of a disc-type oil collection mechanism according to an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of an autonomous path planning and obstacle avoidance process according to an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the architecture of an oil spill identification model for water surfaces according to an embodiment of the present invention;

[0066] Figure 5 This is a schematic diagram of the architecture of a multi-layer feature extraction module according to an embodiment of the present invention;

[0067] Figure 6 This is a schematic diagram of the loss convergence during the training of an oil spill recognition model according to an embodiment of the present invention;

[0068] The markings in the diagram are: 1-Unmanned surface patrol vehicle, 2-Disc refueling mechanism, 3-Refueling tank, 4-High-definition camera, 5-Data transmission antenna, 6-Control cabin, 7-Automatic positioning and navigation sensor, 8-Obstacle avoidance radar. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0070] like Figures 1 to 6 As shown, in a first aspect, this embodiment discloses an automatic oil spill recovery system for water surfaces. The system includes an unmanned surface patrol vehicle 1 and an oil recovery device, an image acquisition device, an automatic positioning and navigation subsystem, and an automatic control subsystem installed on the unmanned surface patrol vehicle 1.

[0071] The image acquisition device acquires water surface images in real time and sends the water surface images to the automatic control subsystem; furthermore, the water surface images carry the shooting location information of the unmanned water surface cruiser 1.

[0072] The automatic control subsystem identifies the oil spill area based on the water surface image and instructs the automatic positioning and navigation subsystem to obtain the location information of the oil spill area;

[0073] The surface unmanned patrol vehicle 1 arrives at the oil spill area based on the location information and uses the oil recovery device to perform the oil spill recovery task.

[0074] In the above technical solution, the unmanned surface patrol vehicle 1 is used to patrol within the target water surface area (i.e., the pre-designated water surface area to be patrolled), and accurately perceives the position of the unmanned surface patrol vehicle itself through the automatic positioning and navigation subsystem.

[0075] The image acquisition device is used to acquire water surface images in real time during the cruise of the unmanned surface patrol vehicle 1, transmit the water surface images to the automatic control subsystem, and obtain the actual position information corresponding to the water surface images through the automatic positioning and navigation subsystem. In this embodiment, during the real-time acquisition of water surface images, the actual position information (i.e., coordinate positioning) of the water surface images is also recorded and saved, that is, the water surface images carry the shooting position information of the unmanned surface patrol vehicle 1.

[0076] The automatic control subsystem described in this embodiment integrates an intelligent oil-water image recognition module. The intelligent oil-water image recognition module is used to accurately identify oily areas on the water surface image and transmit the water surface image with oily areas to the automatic positioning and navigation subsystem.

[0077] The automatic positioning and navigation subsystem is used to locate the actual location of the oil spill area in the water surface image based on the actual location information corresponding to the water surface image, and then automatically plan the optimal path for the unmanned surface patrol vehicle 1 to recover the oil spill, and guide the unmanned surface patrol vehicle 1 to the water surface area with the oil spill area.

[0078] After the unmanned surface patrol vehicle 1 reaches the water surface area with oil spills, it automatically performs the oil spill recovery task through the oil recovery device.

[0079] The aforementioned technical solution is an intelligent oil spill recovery system specifically designed for surface crude oil cleanup and recovery. It effectively overcomes the shortcomings of traditional technologies, aiming to revolutionize the current state of surface crude oil cleanup and recovery (i.e., surface oil spill recovery). This solution comprehensively improves the safety, efficiency, and economic benefits of crude oil collection, provides more reliable personal safety guarantees for operators, and optimizes the oil recovery process, achieving a dual breakthrough in cost control and efficiency improvement. Adopting this technical solution not only enhances the automation level of the oil recovery equipment but also significantly improves its operational accuracy and efficiency in complex aquatic environments, contributing technological strength to water environmental protection and crude oil recovery.

[0080] In this embodiment, the surface unmanned cruiser 1 is a microboat. The oil retrieval device includes an oil retrieval mechanism and an oil retrieval tank 3. The oil retrieval mechanism collects oil slicks from the water surface and collects them into the oil retrieval tank 3. The automatic positioning and navigation subsystem includes an automatic positioning and navigation sensor 7 and an obstacle avoidance radar 8. The surface unmanned cruiser 1 accurately perceives its own position through the automatic positioning and navigation sensor 7 and avoids obstacles through the obstacle avoidance radar 8. The image acquisition device includes a high-definition camera 4, which acquires real-time images of the water surface. In specific implementation, the oil retrieval mechanism is a disc-type oil retrieval mechanism 2, that is, a traditional disc-type oil retrieval mechanism 2 is deployed on the microboat. The disc-type oil retrieval mechanism 2 is as follows: Figure 2 As shown, this belongs to existing technology and will not be elaborated upon here.

[0081] Specifically, to significantly improve the operational efficiency, accuracy, and automation of the traditional disc-type oil skimmer 2 in aquatic environments, this embodiment integrates the disc-type oil skimmer 2, a high-definition camera 4, and an automatic positioning and navigation sensor 7 onto the miniature vessel. The high-resolution camera mounted on the miniature vessel enables real-time monitoring and recording of the water surface conditions, facilitating the identification and detection of oil spills.

[0082] Furthermore, combined with advanced automatic positioning and navigation sensors 7, the aforementioned automatic oil spill recovery system ensures that the micro-vessel accurately locates itself in complex aquatic environments, effectively avoids obstacles, and autonomously navigates to the target area (i.e., the water surface area with oil spills), greatly enhancing the flexibility and safety of operations. This integrated vision and positioning technology not only optimizes the execution process of waterborne operations but also provides strong support for subsequent data analysis and decision support.

[0083] The hull structure of the miniature boat in this embodiment is as follows: Figure 1As shown, the entire hardware system of this embodiment is based on the micro surface unmanned cruiser 1 (i.e., micro boat), and integrates an oil recovery device (mainly including a disc oil recovery mechanism 2 and an oil recovery tank 3), an image acquisition device (mainly including a high-definition camera 4 and a data transmission antenna 5), ​​an automatic positioning and navigation subsystem (including an automatic positioning and navigation sensor 7 and an obstacle avoidance radar 8 set outside the control cabin 6) and an automatic control subsystem (mainly including an intelligent oil-water image recognition module), etc. The surface unmanned cruiser 1 is also equipped with an automated watertight control cabin 6 and a power supply system. The automatic positioning and navigation subsystem and the automatic control subsystem are both set inside the automated watertight control cabin 6.

[0084] Unlike existing image acquisition platforms (such as drones), this embodiment selects a novel platform—a surface unmanned patrol vehicle 1 (i.e., a micro-unmanned vessel)—to carry the image acquisition device. This platform not only provides a different observation perspective than previous platforms but also empowers the system to directly participate in surface operations, namely, collecting oil spills from the water surface through an oil collection device on the micro-vessel. For the different perspectives and environmental conditions (application scenarios) brought about by the novel platform (i.e., the micro-unmanned vessel), this embodiment adopts a new image recognition strategy, specifically implemented through a surface oil spill recognition model. The surface oil spill recognition model in this embodiment employs a multi-layer feature extraction module and a graph convolutional neural network, and embeds an attention mechanism module within the model. The attention mechanism dynamically adjusts the model's focus on input information, significantly improving the model's efficiency and performance when processing complex images. Especially in image recognition and understanding tasks, it helps the model focus more on key information, thereby making more accurate judgments. The aforementioned technical solution, by effectively combining the unmanned surface patrol vehicle 1 with the oil spill identification model, not only achieves comprehensive monitoring of oil pollution on the water surface but also provides a practical solution for subsequent oil spill cleanup, demonstrating its enormous potential in the field of marine environmental protection. It should be noted that the platform on which the image acquisition device is mounted determines the viewing angle and environmental conditions for image capture, which in turn affects the implementation details and algorithm design of subsequent solutions, and also influences the specific development of the technical approach.

[0085] In this embodiment, the miniature unmanned surface patrol vehicle 1 captures images of the water surface in the patrol area using a high-definition camera 4. A smart oil-water image recognition module (which integrates an "intelligent oil-water recognition algorithm" classifier, i.e., a water surface oil spill recognition model) distinguishes between water and oil, thereby identifying oily areas on the water surface. After identifying the oily areas, the automatic positioning and navigation subsystem onboard the miniature unmanned surface patrol vehicle 1 performs "autonomous path planning and obstacle avoidance," the process of which can be referenced. Figure 3Upon reaching the oil spill area, the miniature surface unmanned vehicle 1 collects the oil spill from the water surface using the disc-type oil collection mechanism 2 mounted at its front end, collecting the oil spill in the water into the oil collection tank 3 inside the unmanned vehicle, thus completing the entire surface oil spill collection operation.

[0086] In simple terms, autonomous path planning and obstacle avoidance can be understood as follows: Several points are planned in the host computer of the micro-unmanned surface vessel (USV), for example, five points. The goal is for the USV to navigate sequentially from point 1 to point 5. After the path planning is completed, it is sent to the USV for execution. The USV navigates sequentially from point 1 to point 5. During navigation, if the lidar (i.e., obstacle avoidance radar 8) detects a large obstacle, it executes the obstacle avoidance procedure. After obstacle avoidance, it automatically returns to the initially planned path to continue execution. During navigation, the high-definition camera 4 and the plankton imager monitor the plankton situation on and under the sea surface and send the data to the shore-based system. Further details on autonomous path planning and obstacle avoidance can be found in existing technologies, which will not be elaborated upon here.

[0087] In this embodiment, the automatic positioning and navigation subsystem and the automatic control subsystem together form an advanced automatic control and autonomous navigation system. Data interaction exists between the automatic positioning and navigation subsystem and the automatic control subsystem. Specifically, the control of the micro unmanned vessel is achieved by two parts (i.e., a navigation controller and a motion controller). In this embodiment, the motion controller is integrated into the automatic control subsystem, and the navigation controller is integrated into the automatic positioning and navigation subsystem. Data interaction exists between the navigation controller and the motion controller.

[0088] The motion controller is mainly responsible for the underlying control of actuators such as propellers, data information and control of sensors such as batteries and range extenders, and data communication with shore-based systems; in layman's terms, this part controls the ship's left, right, forward, backward, acceleration and deceleration.

[0089] The navigation controller is mainly responsible for data acquisition from high-bandwidth sensors such as inertial navigation and lidar. It has an internal ROS system that ultimately obtains the control signals from the two underwater thrusters and sends them to the motion controller for execution. In layman's terms, this part realizes path planning through system scheduling, which is the planning and control of the ship's motion path and obstacle avoidance control.

[0090] Secondly, the present invention also provides a water surface oil spill identification model, which is applied to the aforementioned automatic water surface oil spill recovery system. The automatic control subsystem integrates an intelligent oil-water image recognition module, which integrates a water surface oil spill identification model (hereinafter referred to as the model). The automatic control subsystem uses the water surface oil spill identification model to identify oily areas based on the water surface image. The water surface oil spill identification model algorithm can accurately analyze the image and identify oily areas on the water surface, providing target guidance for subsequent oil recovery operations.

[0091] The oil spill identification model includes:

[0092] A multi-layer feature extraction module is used to extract visual features and capture oil slick features from the input water surface image (hereinafter referred to as the image), as well as extract spatial information of salient regions. The visual features include local details and global background. This module is responsible for extracting oil-water related features, which are subsequently used for model training and prediction. The module is designed to ensure that it can capture the visual information most helpful in distinguishing between oil and water. Visual features refer to the features of the digital image captured by the high-definition camera at the pixel scale (a small square of pixels, which can be understood as a small grid in a photo), such as the color, texture, edges, and shape within the grid. Oil slick features refer to physically visible and measurable features, such as the color of the oil film, the shape of the oil film, the color change at the boundary between the oil film and water, and the internal tension of the oil film causing it to be planar rather than three-dimensional. Visual features can reflect all the information in the water surface image; oil slick features refer to visual features related to oil slicks, meaning that oil slick features can be understood as a subset of visual features.

[0093] The Graph Convolutional Neural Network (GCN) module is used to enrich the contextual information in the visual features of each salient region and extract the contextual information between multiple salient regions. When processing features and regions, the GCN considers not only individual features but also the relationships between features; and not only individual regions but also the relationships between regions. This helps the model understand how features interact in space and how regions interact in space, thus enabling better identification of oil and water.

[0094] The attention mechanism module assigns weights to visual features based on their correlation with the oil-water relationship, highlighting the most indicative areas in the water surface image. This allows the oil spill identification model to focus on the most relevant visual features for prediction, thereby identifying oil slicks and their locations in the water surface image. Here, visual features refer to useful information extracted from the water surface image. This information represents certain visual patterns or attributes in the image, such as color, texture, and shape, and may also include the reflection of oil slicks, color variations, and the texture of the oil film.

[0095] By combining three components—a multi-layer feature extraction module, a graph convolutional neural network module, and an attention mechanism module—the proposed oil spill identification model provides a comprehensive and accurate method for detecting oil pollution on water surfaces, maintaining high precision and accuracy.

[0096] 1. Multi-layer feature extraction module

[0097] The multi-layer feature extraction module, serving as the backbone network for oil spill detection, extracts visual features and spatial information of salient regions; it forms the foundation of the oil spill recognition model. It utilizes the multi-layer structure of deep convolutional neural networks (CNNs) to extract rich visual features from the input image. These features capture not only local details but also global background, making them ideal for identifying various types of oil spills on the water surface. Furthermore, the network preserves spatial information, which is crucial for accurately locating oil spills in the image.

[0098] The multi-layer feature extraction module extracts image features through multiple layers of deep convolutional neural networks (CNNs). Lower layers may capture basic visual elements such as edges and textures, while higher layers may capture more complex patterns, such as the overall shape and distribution of oil stains. These features provide rich information for subsequent analysis.

[0099] As a preferred technical solution, the multi-layer feature extraction module includes:

[0100] The feature extraction network, based on a deep convolutional neural network, is used to extract visual features and capture oil slick features from the input water surface image and extract feature maps.

[0101] The RPN module is used to dynamically generate multiple anchor points on the feature map, predict the probability score of the anchor point belonging to the oil pollution on the water surface and the position offset relative to the anchor point for each anchor point, and then propose a set of high-probability candidate regions based on the predicted probability score and position offset. Each candidate region is a suggestion, and a preliminary binary classification is performed on each suggestion to determine whether there is oil pollution and its approximate location. The candidate regions with oil pollution are designated as salient regions.

[0102] In this embodiment, candidate regions refer to areas in a water surface image that are initially identified by the oil spill detection model algorithm as potentially containing oil contamination. These regions are obtained based on image feature extraction and analysis, such as visual features like color, texture, and shape, which may match known oil contamination characteristics. In the oil spill detection model algorithm, candidate regions are the output of the Region Proposal Network (RPN), serving as the basis for further analysis and refinement. The RPN predicts possible oil contamination locations by setting anchor points on the feature map; these anchor points are scored and adjusted based on their degree of matching with actual oil contamination.

[0103] The relationship between candidate regions and oil spill features lies in the fact that candidate regions are predicted by the oil spill identification model algorithm based on visual features extracted from images. These visual features are related to the physical and chemical properties of oil spills. The algorithm learns, through training, which visual features are highly correlated with the presence of oil spills, thus enabling it to identify potential oil spill areas in new images. This method identifies image blocks containing oil spills and further merges these blocks. This embodiment employs a Graph Convolutional Neural Network (GCN), which effectively reduces computational load while better distinguishing oil spills and pinpointing their specific locations.

[0104] The RoI pooling module is used to extract RoI features from the feature map, adjust these RoI features of different sizes, convert them into feature sub-maps of uniform and preset size, and map the feature sub-maps back to the entire feature map.

[0105] The R-CNN module outputs the entire feature map obtained by the RoI pooling module to its own fully connected layer to perform oil spill prediction and location regression. Oil spill prediction identifies whether there is oil in the water surface image, and location regression predicts and locates the oil spill's position within the image. In this embodiment, oil spill prediction (identifying the presence of oil) and location regression (predicting the location of the oil spill area) are performed simultaneously using a single algorithm. The aforementioned RPN module only identifies anomalies in the image.

[0106] In this embodiment, the multi-layer feature extraction module serves as the key architecture, consisting of four main components: a feature extraction network, an RPN module (Region Proposal Network or Region Generation Network), an RoI pooling module, and an R-CNN module. The core task of the multi-layer feature extraction module is to generate and optimize candidate regions. Initial classification and localization are achieved through the RPN module. Subsequent stages refine these candidate regions using a graph convolutional neural network module, followed by processing with an attention mechanism module, ultimately achieving accurate classification and detail enhancement of oil stain features.

[0107] This embodiment innovatively introduces a global contextual attention mechanism (i.e., a global attention mechanism model) into the feature extraction network. First, it builds bridges between feature maps of adjacent resolutions to ensure information continuity. Then, it extracts global contextual information from the low-resolution feature map to enhance the weights of high-resolution features, thereby accurately depicting the spatial layout of class pixels. This strategy not only solidifies the expressiveness of high-level features but also ensures higher classification accuracy while maintaining computational efficiency. The global contextual attention mechanism is introduced here to reduce the complexity of the algorithm. For example, in an oil-water image, many features (e.g., 100 features) can identify an object, but calculating all 100 features would be computationally too intensive and slow, failing to meet real-time requirements. Therefore, an attention mechanism is introduced, focusing on only 5 features to prove the presence of oil on the water surface, identifying the oil-slicked area. This significantly reduces computational load while enabling accurate identification.

[0108] The RPN module (i.e., the RPN network) cleverly maps the generated candidate regions onto the feature maps produced by the convolutional network through an "anchor point" mechanism, achieving seamless integration between the two and significantly improving the speed and accuracy of object detection. Anchor points with specific sizes and aspect ratios are set at various locations on the output feature map of the RPN network to match actual oil stains, enabling precise classification and fine-tuning of oil stain locations.

[0109] like Figure 5 As shown, in the multi-layer feature extraction process, preliminary localization and classification suggestions are first given, and their accuracy directly affects the error rate of subsequent re-examination. The RoI (Region of Interest) is generated by filtering the RPN prediction anchor points. It is obtained by filtering a large number of candidate regions during the training phase, and is directly used as the RoI during the testing phase.

[0110] The RPN module has a five-level subsystem embedded in it.

[0111] (1) Anchor point generation: For each coordinate point on the feature map, nine anchor points are generated, covering three areas and three aspect ratios, aiming to fully cover potential oil pollution areas.

[0112] (2) RPN convolutional network: The above anchor points are processed by a convolutional network to obtain the prediction score and displacement value.

[0113] (3) Calculate RPN loss: This is performed only during the training phase. The anchor point is compared with the label to distinguish between positive and negative samples. The loss function is calculated by combining the predicted score and the displacement value.

[0114] (4) Recommendation generation: Select high-quality anchor points from the output of the RPN convolutional network to provide a preferred candidate set for subsequent networks.

[0115] (5) Proposal screening and RoI acquisition: Select the final RoI from the suggestions screened in step four for subsequent deep learning model analysis.

[0116] This series of processes ensures that the most relevant information is captured from multi-scale features, providing a solid foundation for oil spill detection.

[0117] 1.1 Feature Extraction Network

[0118] The feature extraction network is the cornerstone of the entire model architecture, performing exceptionally well in low-contrast images and efficiently extracting clear feature maps. Therefore, this embodiment employs a 13-layer convolutional network similar to the VGG16 architecture, specifically designed to magnify images and capture oil stain features. This network utilizes 3×3 convolutional kernels; while these small kernels have a limited number of parameters, their deep hierarchical structure endows the network with powerful non-linear expressive capabilities, thereby enhancing its efficiency in learning and simulating computational intelligence and neuroscience principles.

[0119] This network architecture consists of five core convolutional modules, each outputting feature maps of different dimensions: 64, 128, 256, 512, and 512 channels respectively. This design enables the network to capture image information at different levels of abstraction, from macroscopic contours to minute textures, leaving no detail overlooked and laying a solid foundation for subsequent feature analysis and object recognition. Through this series of carefully designed convolutional operations, even under challenges of insufficient image contrast, the extracted features are ensured to possess high discriminative power and expressiveness, thereby driving improvements in overall detection performance.

[0120] 1.2 RPN Module

[0121] RPN (Region Proposal Network) is a network used to generate candidate regions, primarily used in Faster R-CNN. Its main function is to generate candidate bounding boxes (region proposals) that may contain objects, which are then used for further object detection.

[0122] The workflow of an RPN network can be divided into the following steps:

[0123] Input image processing: The input image is first processed by a convolutional neural network (such as ZF or VGG) to generate a feature map. Feature map processing: The feature map is convolved with a 3x3 convolutional kernel to generate a new feature map.

[0124] Anchors are generated for each point on the feature map, and each anchor corresponds to a region in the original image.

[0125] Classification branch: Classify each anchor point to determine whether it contains an object.

[0126] Border regression branch: Perform border regression on the anchor points containing the object, and adjust their position and size.

[0127] In this embodiment, the RPN module dynamically generates a series of anchor points on the feature map. For each anchor point, it accurately evaluates and predicts its probability score of belonging to oil sludge on the water surface and its positional offset relative to the anchor point. Based on these predictions, the RPN module can propose a set of high-probability candidate regions, the so-called "suggestions," and perform preliminary binary classification on each suggestion to determine whether oil sludge exists and its approximate location. Notably, the RPN module and the subsequent R-CNN detection network share the convolutional layers of the entire feature map. This design saves computational resources and ensures that the R-CNN can receive high-quality region suggestions, thereby significantly improving the detection accuracy of the overall model. At this stage, the output feature map contains 512 channels, which are further refined to 18 and 36 channels in the subsequent classification and localization branches, respectively, for more refined processing.

[0128] 1.3 RoI Pooling Module

[0129] In the field of deep learning, Region of Interest (RoI) is mainly used for object detection. Its introduction into the Fast R-CNN framework aims to improve the speed and efficiency of the R-CNN series of algorithms.

[0130] The main functions of RoI Pooling are: (1) Size normalization: The RoI Pooling layer receives candidate regions (also known as proposal boxes or region proposals) of different sizes and proportions and converts them into feature vectors of a fixed size, which makes these feature vectors available for processing by fully connected layers in the network. (2) Improved computational efficiency: Compared with the previous R-CNN method, it avoids multiple forward passes on the entire image and instead performs a single forward pass on the shared convolutional feature map, and then applies RoI Pooling to each candidate region.

[0131] In this embodiment, given that the later-stage R-CNN module relies on fully connected layers for deep learning processing, RoI pooling is crucial to ensure that all oil slick feature sub-images input to this layer have a uniform size (preferably 7×7 in this embodiment). Faced with oil slick images of varying areas on the water surface, the corresponding RoIs naturally exhibit different representations on the feature maps. The RoI pooling mechanism intelligently adjusts these RoI features of different sizes, converting them into a uniform feature map of a preset size, ensuring that it adapts to the needs of detecting oil slicks on the water surface regardless of the original size of the oil slick image. During this process, the number of channels in the feature map remains constant, effectively maintaining data consistency and integrity.

[0132] Through the collaborative work of the RPN and RoI pooling modules, the entire system not only effectively filters high-quality oil spill candidate regions from the original image, but also ensures that these candidate regions can enter the subsequent R-CNN detection network in a standardized form for in-depth feature analysis and classification, ultimately achieving accurate identification and localization of oil spills on the water surface. This design not only greatly optimizes computational efficiency but also significantly improves the model's detection performance in complex environments.

[0133] 1.4 R-CNN Module

[0134] The region of interest (RoI) feature map obtained by RoI pooling is mapped back to the entire feature map and output to the fully connected network of the R-CNN module to perform oil spill prediction and location regression. This fully connected layer outputs 2048 feature maps. The visual features of the i-th anchor point output by Faster R-CNN are... And spatial information is S i =[x i1 y i1 x i2 y i2 ], where m represents the number of significant regions.

[0135] 2. Graph Convolutional Neural Network Module

[0136] The Graph Convolutional Neural Network (GCN) module enriches the contextual information in the visual features of each salient region to obtain the graph structure of the features, and extracts the contextual information between multiple salient regions to obtain the graph structure of the region. After the initial feature extraction, the oil spill identification model uses GCN to further enrich the contextual information of the visual features and extract the contextual information between multiple salient regions. GCN is designed to process graph data structures, where nodes represent objects and edges represent the relationships between them.

[0137] To enhance the contextual information in the visual features of each region, this embodiment uses GCN to update the object representation. That is, GCN enriches the contextual information in the visual features of each salient region to obtain the graph structure of the features, such as... Figure 4 The left side of the dashed box in the image shows the convolutional neural network. In the graph structure of the features, nodes represent visual features, and edges represent the relationships between them. By processing the extracted features through the graph convolutional network, the model can learn the spatial relationships and contextual information (i.e., interactions) between different features, thereby gaining a deeper understanding of the specific environment surrounding the oil spill, such as the relative position of the oil spill to other objects, thus improving its ability to understand oil spills.

[0138] After obtaining the visual features and spatial information of salient regions using Faster R-CNN, this embodiment uses GCN to update the object representation. Specifically, GCN extracts the contextual information between multiple salient regions to obtain the graph structure of the regions, such as... Figure 4 The right side of the dashed box in the middle of the diagram shows the convolutional neural network. First, a fully connected graph is constructed, and then further refined using precise contextual information between multiple salient regions to obtain a spatial graph network (i.e., the graph structure of the regions).

[0139] In a graph structure of regions, nodes represent regions, and edges represent the relationships between them. That is, each object region in the image is considered a node (vertex), and a fully connected indirect graph is obtained by constructing a relationship graph, such as... Figure 4 The right side of the dashed box in the Chinese image of the convolutional neural network is shown in the upper part of the image. Each edge represents the relationship between two regions. The spatial information representing the location of a region is a four-dimensional spatial vector S. i =[x i1 y i1 x i2 y i2 ], where (x i1 y i1 ) represents the coordinates of the top-left corner of the bounding box, while (x) represents the coordinates of the top-left i2 y i2 ) represents the coordinates of the bottom right corner of the bounding box.

[0140] The identification of inter-region correlations is carried out according to the following steps: (1) The visual features of the two regions are input into a multilayer perceptron to obtain feature fusion, and then the corresponding elements of the two feature embeddings are multiplied to obtain a correlation score. (2) The correlation between the two regions is determined based on the size of the overlapping area. If the two regions have a large overlapping area, it means that there is a strong correlation between the two regions. If the two regions have no overlapping part, it is considered that the correlation between the two regions is weak, which means that there is no edge connecting the two nodes. In addition, different regional relationship categories are identified, such as interior, coverage and overlap. Based on spatial relationships, some irrelevant regional relationships are removed from the fully connected graph to obtain a relatively sparse graph, such as Figure 4 The right side of the dashed box in the middle of the image is shown in the figure below.

[0141] To enhance the contextual information in the visual features of each region, we use GCN to update the object representation. If the image contains m salient regions, as nodes, we use an m×m adjacency matrix A to represent the graph structure, where A is defined if there is an overlapping region between node i and node j. ij =1, otherwise A ij =0. Given a target node i and its neighboring nodes j ∈ N(i) in an image, where N(i) is the set of neighboring nodes of node i, and the visual feature representations of nodes i and j are h, respectively. i and h j To obtain the correlation score s between nodes i and j. ij By splicing h i and h j To capture the visual features, we first train a fully connected layer:

[0142]

[0143] Where ω a and W a These are the learning parameters, σ is the non-linear activation function, and [h] is the [learning parameters]. i h j The symbol ] indicates a concatenation operation. We use the correlation score s... ij The softmax function is used to obtain the weights α. ij ,like Figure 4 The right side of the dashed box in the middle of the image is shown in the figure below.

[0144]

[0145] For graph convolution, first consider the neighboring nodes h i , j∈N(i) through the learned linear transformation W b These transformed representations are weighted by α. ijThe aggregated features are then used to obtain the updated node features h through the activation function σ. vi :

[0146]

[0147] The output feature of node i in the last layer of GCN is Hi, and the set of features of all nodes is H.

[0148] In this embodiment, Figure 4 The left side of the dashed box in the middle of the image illustrates how to construct a graph structure of features from the extracted features that helps GCN work. Figure 4 The right side of the dashed box in the middle of the image illustrates how to construct a graph structure from the acquired regions that facilitates the operation of the GCN. Figure 4 The overall presentation demonstrates how the graph structure described above can be used to enhance feature and region representations, further supporting the effectiveness of the proposed surface oil spill detection method (i.e., identification through a surface oil spill identification model) in this embodiment.

[0149] 3. Attention Mechanism Module

[0150] The attention mechanism module is primarily used to explore the semantic association between visual features and oil-water. To optimize the oil-water classification process, the oil spill recognition model in this embodiment integrates an attention mechanism. This allows the model to focus on the most relevant features for prediction, determining whether an image contains oil or water. The attention mechanism assigns weights to features based on their relevance to specific oil-water pairs, effectively emphasizing regions crucial to the final decision. This not only improves classification accuracy but also makes the model's decision-making process more interpretable by highlighting the most indicative areas in the image.

[0151] The application of attention mechanisms helps models better focus on key features of oil spills, thereby improving the accuracy and efficiency of identifying different types of oil spills. This mechanism is particularly crucial in complex aquatic environments, as it helps models extract visual features closely related to oil spills from a large amount of background information, reducing false positives and false negatives, and improving overall detection performance. First, for the region visual representation H obtained by graph neural network convolution, a self-attention mechanism is used to update the feature set of each node to obtain a new feature set.

[0152]

[0153] Where H T W is the transpose of H, and d is the dimension of H. To obtain a visual representation related to oil slicks on the water surface, we utilize a learnable parameter matrix W. m W m Used as a guiding matrix to adjust visual representation W m and The similarity score between them is calculated as follows:

[0154]

[0155] For the i-th region, the score r is normalized using the softmax function. i To obtain the probability of each area being an oil spill:

[0156]

[0157] By embedding an attention mechanism module into the network architecture, the model's operating efficiency and recognition performance can be significantly improved, especially when processing complex images, such as image recognition and understanding. The attention mechanism enables the model to more accurately focus on key features in the image and effectively filter out irrelevant information, thereby greatly improving processing speed and analysis accuracy while ensuring accuracy.

[0158] Thirdly, the present invention also provides a method for training a water surface oil spill recognition model, for training the aforementioned water surface oil spill recognition model, the training method comprising:

[0159] Acquire online images of oil spills on the water surface;

[0160] An image dataset is composed of oil-water samples selected from the online images, the oil-water samples including various oil-water types;

[0161] The image dataset is divided into a training sample set and a test sample set, with the ratio of the training sample set to the test sample set set set to 4:1. That is, 80% of the entire image dataset is randomly selected as the training sample set and the other 20% is selected as the test sample set.

[0162] The oil-water samples in the training sample set are labeled;

[0163] The labeled oil-water samples are input into a pre-built oil spill identification model for training.

[0164] The optimized oil spill recognition model was obtained after training.

[0165] The training method for the above-mentioned oil spill recognition model will be further explained below.

[0166] 1. Construct an image dataset

[0167] By setting up a test bench and using a line scan camera to acquire online images of oil pollution on the water surface, a total of 5,000 "oil-water" samples were obtained after screening. These samples were then used to form an image dataset. In this embodiment, the oil-water samples include 5 types of oil-water samples as an example. There are 1,000 samples of each "oil-water" type, of which 4,000 are used as training samples (i.e., forming the training sample set) and 1,000 are used as test samples (i.e., forming the test sample set).

[0168] The oil-water samples in the training sample set are labeled, and the "oil-water" samples are labeled using rectangular boxes through labeling software.

[0169] 2. Training conditions

[0170] In this preferred embodiment, the CPU is an Intel Xeon W-2135, the operating system is Ubuntu 18.04, the RAM is 32GB, the GPU is a GeForce RTX 2080Ti with 8GB of video memory, the Python version is 3.6.9, the CUDA version is 10.0, and the CUDNN version is 7.4.1.

[0171] 3. Test setup and evaluation metrics

[0172] To further demonstrate the effectiveness of the proposed method, the pre-built oil spill identification model was trained using the aforementioned dataset (i.e., the training sample set). The specific training method is as follows: the initial learning rate of the model was set to 0.0001; the Adam optimizer was used; the learning rate was decayed every 5 epochs at a rate of 0.1; the batch size was set to 8 (the batch size is the number of samples selected in one training epoch; due to the GPU limitations of the device, the batch size was chosen to obtain the best optimization and the highest training speed). Figure 6 The diagram shows the loss value of the proposed model trained for 100 epochs on the image dataset of this embodiment, where the vertical axis represents the loss value and the horizontal axis represents the number of training iterations. During training, the loss value converged rapidly and was eventually set at approximately 0.18. The above steps allow for the optimization of the trained oil spill recognition model.

[0173] In this embodiment, the oil spill detection model is preferably trained using a multi-task loss function, combining classification loss and edge localization regression loss for unified training, ultimately outputting the corresponding classification results and edge location information. This approach improves detection accuracy, and is particularly suitable for detecting small-target oil spills.

[0174] This loss function takes into account both classification and localization tasks, and its formula is as follows:

[0175]

[0176] in,

[0177] This represents the probability that the i-th anchor point predicted by the model is the target.

[0178] r * i This represents the true category label of the i-th anchor point, which is 1 if a target exists and 0 otherwise.

[0179] S i and S i * Let represent the regression predicted value and the actual value of the position of the i-th anchor point, respectively.

[0180] N cls and N reg These are the normalized values ​​for the classification and regression tasks, used to balance the loss contributions of the two tasks.

[0181] L cls and L reg These are the classification loss and regression loss functions, respectively.

[0182] p i * It is a binary variable that is 1 when the i-th anchor is a positive example and 0 otherwise. It is used to control that only positive examples participate in the calculation of regression loss.

[0183] λ is a hyperparameter used to adjust the weighting between classification loss and regression loss.

[0184] In this way, the model not only learns how to accurately classify oil and water, but also how to precisely locate the edges of oil spills. This is particularly important for detecting small-target oil spills, as these targets occupy a small portion of the image and are easily affected by background noise. Accurate edge localization helps the model better identify and distinguish targets. By optimizing this multi-task loss function, the model can improve its performance in detecting small-target oil spills in complex environments, thereby achieving higher detection accuracy and robustness. This has significant implications for environmental protection and the prevention of oil spills on water surfaces.

[0185] The classification loss function L in this embodiment cls Preferred options are: in, r represents the probability that the i-th anchor point predicted by the model is the target. i *This represents the true class label of the i-th anchor point (1 if a target is present, 0 otherwise). The article does not explicitly specify the form of the classification loss function used, but commonly used classification loss functions include cross-entropy loss, which is suitable for binary or multi-class classification problems. Further details on classification loss functions can be found in existing techniques and will not be elaborated upon here.

[0186] The regression loss function L in this embodiment reg Preferred options are: Among them, s i Let s represent the regression prediction value of the position of the i-th anchor point. i * This represents the true value. Specific regression loss functions include, but are not limited to, commonly used regression loss functions, such as Mean Squared Error (MSE) or Mean Absolute Error (MAE). These functions measure the difference between the predicted and true values. Further details on regression loss functions can be found in existing techniques and will not be elaborated upon here.

[0187] Fourthly, the present invention also provides a method for identifying oil spills on water surfaces, based on the aforementioned oil spill identification model, the method comprising:

[0188] Acquire images of the target water surface area as images to be identified;

[0189] The image to be identified is input into the trained oil spill recognition model, and the recognition information of the image to be identified is output. The recognition information includes classification information and location information. The classification information is whether there is oil in the image to be identified, and the location information is the specific location of the oil in the image to be identified.

[0190] In this embodiment, the automatic control subsystem integrates an intelligent oil-water image recognition module, which incorporates a water surface oil spill recognition model. This module accurately identifies oil-contaminated areas on the water surface image (i.e., obtains recognition information) using the model. Based on this recognition information, it acquires an image of the water surface containing the oil-contaminated area and transmits it to the automatic positioning and navigation subsystem. Upon receiving the information about the oil-contaminated area (i.e., the water surface image containing the oil-contaminated area), the automatic positioning and navigation subsystem automatically plans the optimal path and guides the micro-surface unmanned vehicle directly to the target area (i.e., the water surface area containing the oil-contaminated area). Upon arrival, the disc oil skimmer autonomously initiates the oil skimming process, efficiently completing the oil recovery task.

[0191] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs, characterized in that, when the one or more programs are executed, they can implement the aforementioned oil spill identification model training method or the aforementioned oil spill identification method.

[0192] The parts not covered in this embodiment are the same as or can be implemented using existing technologies, and will not be further described here.

[0193] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic oil spill recovery system for water surfaces, characterized in that, The system includes an unmanned surface cruiser and an oil recovery device, an image acquisition device, an automatic positioning and navigation subsystem, and an automatic control subsystem installed on the unmanned surface cruiser. The image acquisition device acquires water surface images in real time and sends the water surface images to the automatic control subsystem; The automatic control subsystem identifies the oil spill area based on the water surface image and instructs the automatic positioning and navigation subsystem to obtain the location information of the oil spill area; The unmanned surface patrol vehicle arrives at the oil spill area based on the location information and uses the oil recovery device to perform the oil spill recovery task.

2. The automatic system for recovering oil spill on water surface according to claim 1, characterized in that, The water surface image carries the location information captured by the unmanned surface patrol vehicle.

3. A water surface oil spill identification model applied to the water surface oil spill automatic recovery system of any one of claims 1-2, characterized in that, The automatic control subsystem integrates the oil spill detection model and uses the oil spill detection model to identify oil-contaminated areas based on the water surface image. The oil spill detection model includes: A multi-layer feature extraction module is used to extract visual features from the input water surface image, capture oil slick features, and extract spatial information of significant areas. The visual features include local details and global background. The graph convolutional neural network module is used to enrich the contextual information in the visual features of each salient region and to extract the contextual information between multiple salient regions. The attention mechanism module is used to assign weights to the visual features based on their correlation with oil and water, highlighting the most indicative areas in the water surface image, so that the oil spill recognition model can focus on the most relevant visual features for prediction, thereby identifying the oil spill in the water surface image and its location.

4. The water surface oil spill identification model of claim 3, wherein, The multi-layer feature extraction module includes: The feature extraction network, based on a deep convolutional neural network, is used to extract visual features and capture oil slick features from the input water surface image and extract feature maps. The RPN module is used to dynamically generate multiple anchor points on the feature map, predict the probability score of the anchor point belonging to the oil pollution on the water surface and the position offset relative to the anchor point for each anchor point, and then propose a set of high-probability candidate regions based on the predicted probability score and position offset. Each candidate region is a suggestion, and a preliminary binary classification is performed on each suggestion to determine whether there is oil pollution and its approximate location. The candidate regions with oil pollution are designated as salient regions. The RoI pooling module is used to extract RoI features from the feature map, adjust these RoI features of different sizes, convert them into feature sub-maps of uniform and preset size, and map the feature sub-maps back to the entire feature map. The R-CNN module is used to output the entire feature map obtained by the RoI pooling module to the fully connected layer of the R-CNN module itself to perform oil pollution prediction and location regression. The oil pollution prediction is to identify whether there is oil pollution in the water surface image, and the location regression is to predict and locate the position of the oil pollution in the water surface image.

5. The water surface oil spill identification model of claim 4, wherein, A global attention mechanism is introduced into the feature extraction network. The global attention mechanism includes: building bridges between feature maps of adjacent resolutions and extracting global contextual information from low-resolution feature maps to enhance the weights of high-resolution features.

6. The oil spill identification model according to claim 4, characterized in that, The feature extraction network employs a 13-layer convolutional neural network and utilizes 3×3 convolutional kernels. The feature extraction network consists of five core convolutional modules, each of which sequentially outputs feature maps of different dimensions.

7. The water surface oil spill identification model of claim 4, wherein, The RPN module and the R-CNN module share the convolutional layers of the entire feature map.

8. The water surface oil spill identification model of claim 4, wherein, The oil spill detection model is trained using a multi-task loss function, which comprehensively considers both classification and localization tasks. The formula for this multi-task loss function is as follows: in, Pij represents the probability that the i-th anchor point predicted by the water surface oil spill identification model is targeted; r * i true class label for the ith anchor point, 1 when a target is present, otherwise 0; S i and S i * respectively denote the position regression prediction and ground truth of the ith anchor. N cls and N reg These are the normalized values ​​for the classification and regression tasks, used to balance the loss contributions of the two tasks; L cls and L reg are the classification and regression loss functions, respectively; p i * It is a binary variable that is 1 when the i-th anchor point is a positive example and 0 otherwise. It is used to control that only positive examples participate in the calculation of regression loss. λ is a hyperparameter used to adjust the weighting between classification loss and regression loss. 9.A method for training a water surface oil spill identification model, configured to train the water surface oil spill identification model according to any one of claims 3-8, characterized in that, The training method includes: Acquire online images of oil spills on the water surface; An image dataset is composed of oil-water samples selected from the online images, the oil-water samples including various oil-water types; The image dataset is divided into a training sample set and a test sample set, wherein the ratio of the training sample set to the test sample set is set to 4:

1. The oil-water samples in the training sample set are labeled; The labeled oil-water samples are input into a pre-built oil spill identification model for training. The optimized oil spill recognition model was obtained after training.

10. A method for identifying oil spill on water surface based on the model for identifying oil spill on water surface according to any one of claims 3-8, characterized in that, The method includes: Acquire images of the target water surface area as images to be identified; The image to be identified is input into the trained oil spill recognition model, and the recognition information of the image to be identified is output. The recognition information includes classification information and location information. The classification information is whether there is oil in the image to be identified, and the location information is the specific location of the oil in the image to be identified.

11. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions executable by one or more processors to perform operations comprising: When one or more of these programs are executed, the oil spill identification model training method of claim 9 can be implemented, or the oil spill identification method of claim 10 can be implemented.

Images