Improved YOLOv8s-based oil spill pollution source identification method and system

Abstract

The invention discloses an oil spill pollution source recognition method and system based on improved YOLOv8s, and relates to the field of image recognizing.The method comprises the steps that monitoring data of marine pollution sources are collected, and the marine pollution sources comprise oil spill pollution sources; the types of the monitoring data comprise unmanned aerial vehicle aerial images; a pollution source identification model is constructed on the basis of the monitoring data, the pollution source model comprises an improved YOLOv8s network model, and the improved YOLOv8s network model is formed by fusing an MCA module on the basis of the YOLOv8s network model; and completing identification of the marine oil spill pollution source by using the pollution source model. On the basis of a YOLOv8s model, an MCA module is fused, and the module focuses on extracting position information, so that the capability of extracting effective features of a water surface target is enhanced, and the problems of missing detection and excessive detection caused by object folding and crossing can be improved. In addition, the system is high in anti-interference capability and is not easily affected by complex backgrounds.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to an oil spill pollution source recognition method and system based on improved YOLOv8s. Background Art

[0002] In terms of investigation, evidence collection and responsibility confirmation of marine oil spills, it is impossible to supervise ships that are not equipped with AIS systems or have turned off AIS services, and VTS radar stations can only cover sea areas within a certain range near the port area, and their coverage distance is easily limited. In large-scale waters, satellite remote sensing monitoring has the ability to observe and image the earth on a global scale, and has always been an important means of earth observation in countries around the world. Although satellites play an irreplaceable role in the field of earth monitoring, there are still some defects in the face of complex, changeable and time-sensitive emergency monitoring, such as being limited by orbits and revisit time, low time and space resolution, large imaging field of view, easy blur, and long observation period. The UAV remote sensing monitoring technology has the advantages of strong mobility, strong real-time performance, high resolution, and low operating cost, and can adapt to the timely and refined monitoring requirements of marine emergencies, becoming an effective supplementary tool for satellite remote sensing. That is, in large-scale waters, satellite remote sensing monitoring technology is used to preliminarily identify the type of oil spill pollution source and determine whether the oil spill pollution source is caused by an offshore oil platform or a ship. In small offshore waters, drone remote sensing monitoring is used to finely identify pollution sources from both macro and micro perspectives and determine the specific ship type of the oil spill source. Summary of the invention

[0003] In order to solve the technical problems in the above background, the present invention proposes a method and system which has the characteristics of strong versatility, rapidity, simplicity, and adaptability to complex sea surface background environment.

[0004] To achieve the above purpose, the present invention proposes an oil spill pollution source identification method based on improved YOLOv8s, the steps comprising:

[0005] Collecting monitoring data of marine pollution sources, wherein the marine pollution sources include oil spill pollution sources; the types of monitoring data include drone aerial images;

[0006] Based on the monitoring data, a pollution source identification model is constructed, wherein the pollution source model includes an improved YOLOv8s network model, wherein the improved YOLOv8s network model is a YOLOv8s network model integrated with an MCA module; the introduced MCA module is composed of three parallel branches, wherein the top branch is used to capture the interaction between features in the spatial dimension W; the middle branch is used to capture the interaction between features in the spatial dimension H; and the bottom branch is responsible for capturing the interaction between channels C;

[0007] In the top branch and the middle branch, permutation operations are used to capture the long-range dependencies between the channel dimension and any spatial dimension; finally, the outputs of the three branches are aggregated by calculating the average in the integration stage.

[0008] The pollution source model is used to identify the source of marine oil spill pollution.

[0009] Preferably, the collected monitoring data are sorted and labeled, and the monitoring data include offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles; the collected monitoring data are divided into a training set, a validation set and a test set according to a preset ratio.

[0010] Preferably, the MCA module uses global average and standard deviation pooling to aggregate cross-dimensional feature responses in the squeeze transformation, and uses a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors;

[0011] The MCA module captures local feature interactions in a lightweight manner during excitation transformation, overcoming the trade-off between performance and computational overhead.

[0012] The present invention also provides an oil spill pollution source identification system based on improved YOLOv8s, the system is used to implement the above method, including: a collection module, a construction module and an identification module;

[0013] The acquisition module is used to collect monitoring data of marine pollution sources, wherein the marine pollution sources include oil spill pollution sources; the types of monitoring data include drone aerial images;

[0014] The construction module is used to construct a pollution source identification model based on the monitoring data, and the pollution source model includes an improved YOLOv8s network model, and the improved YOLOv8s network model is a fusion of the MCA module on the basis of the YOLOv8s network model; the MCA module is composed of three parallel branches, wherein the top branch is used to capture the interaction between features in the spatial dimension W; the middle branch is used to capture the interaction between features in the spatial dimension H; the bottom branch is responsible for capturing the interaction between channels C; in the top branch and the middle branch, a permutation operation is used to capture the long-range dependency between the channel dimension and any spatial dimension; finally, in the integration stage, the outputs of the three branches are aggregated by calculating the average;

[0015] The identification module is used to use the pollution source model to complete the identification of the marine oil spill pollution source.

[0016] Preferably, the workflow of the acquisition module includes: organizing and labeling the collected monitoring data, wherein the monitoring data includes offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles; dividing the collected monitoring data into a training set, a validation set and a test set according to a preset ratio.

[0017] Preferably, the MCA module uses global average and standard deviation pooling to aggregate cross-dimensional feature responses in the squeeze transformation, and uses a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors;

[0018] The MCA module captures local feature interactions in a lightweight manner during excitation transformation, overcoming the trade-off between performance and computational overhead.

[0019] The water surface environment may present complex scenes, including various background elements such as wind, waves, and currents. These background elements will introduce a lot of redundant information and interference into the image, making it difficult to separate the target from the complex background.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) Based on the YOLOv8s model, the present invention integrates the MCA module, which focuses on extracting position information, enhances the ability to extract effective features of surface targets, and helps to improve the problems of missed detection and multiple detection caused by object folding and crossing.

[0022] (2) The present invention has strong anti-interference ability and is not easily affected by complex backgrounds such as light intensity and wind, waves, and currents. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0024] Figure 1 It is a schematic diagram of the YOLOv8 structure of an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the improved YOLOv8 structure of an embodiment of the present invention;

[0026] Figure 3 FIG. 4 is a schematic diagram of an MCA module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Embodiment 1

[0030] The method steps of this embodiment are as follows:

[0031] S1. Collect monitoring data on marine pollution sources, including oil spill sources. The types of monitoring data include drone aerial images.

[0032] The drone remote sensing data comes from open source data sets and aerial photography experiments. Seven categories of data sets are collected, organized and annotated, including offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles. The images in these data have various rotation angles and lighting conditions, which helps to train a more robust detection model. After the monitoring data is collected, it is organized and annotated. The data organization process mainly involves image preprocessing. This is because there are often noise, lighting changes, inconsistent image sizes and other problems in the original data, which will affect the accuracy of target recognition. After obtaining the original data set, preprocessing is required to perform image correction, image denoising and image enhancement on the original image data.

[0033] When conducting experiments, the collected 8025 data sets were divided into 6200 training sets, 1130 validation sets, and 695 test sets. This allocation strategy ensures that the model can learn on rich data and verify accuracy and generalization ability on independent data. In order to minimize image distortion without affecting detection accuracy, all images were resized to 640×640 and the original aspect ratio was maintained.

[0034] S2. Based on the monitoring data, a pollution source identification model is constructed. The pollution source model includes an improved YOLOv8s network model. The improved YOLOv8s network model is a YOLOv8s network model that integrates an MCA module.

[0035] Although the recognition algorithm based on convolutional neural network has achieved great success in conventional target detection tasks, the detection effect of surface targets in UAV remote sensing scenarios is not good. The main reason is that in UAV aerial images, the size of objects varies greatly, and the single-layer feature map representation ability of convolutional neural networks is limited. In addition, in aerial images, the background is complex and changeable, and the target object is easily disturbed, which affects the recognition effect. Since the YOLO algorithm has the characteristics of strong versatility, fast, simple, and can adapt to complex sea surface background environments, this embodiment proposes a method for identifying oil spill sources based on improved YOLOv8s for UAV remote sensing image data. The typical representative of the single-stage detection model is YOLOv8. According to practical application experience, the real-time performance of the model is high and the detection accuracy has also reached a high level.

[0036] Among visible light target detection algorithms, the YOLO series of models have high real-time performance and accuracy, and therefore have received widespread attention in the field of target detection. Figure 1 As shown in the figure, the YOLOv8 structure can be divided into the backbone feature extraction network, feature enhancement, detection head and other units. The preprocessing operation of the local network module CSP in this network is appropriately simplified, mainly converted into two convolutions, and there is a concept of multi-stack. At the same time, the C3 structure is replaced and converted into a C2f structure, which can better meet the performance requirements of actual applications. The number of channels is appropriately adjusted based on different scales, which has an important role in promoting network learning and also improves the robustness of the algorithm. The CSP module performs two convolutions during operation and multi-stacks different modules to improve the performance of the network. In actual application, the number of channels of the initial convolution is doubled, and then the result is split in half, which is of great significance to improving the calculation speed of the model. The detection head of YOLOv8 uses two independent branches for object classification and predicted bounding box regression. The predicted bounding box regression adopts the DFL and CIoU structures, which improves the detection accuracy and accelerates the convergence of the model.

[0037] YOLOv8 is one of the basic models of the YOLO series. Compared with the original anchor-based detection method, the anchor free method it adopts has higher detection accuracy and detection speed. However, in a complex water surface environment, the floating features of the surface target are complex and the background is diverse. The use of the YOLOv8 detection algorithm may have disadvantages such as positioning error and insufficient target perception. Therefore, in order to solve the above problems, this embodiment proposes an improved YOLOv8 algorithm. Figure 2As shown, after the feature fusion of the YOLO model, a multi-scale convolution attention module (Multidimensional collaborativeattention, MCA) is set. After the introduction of this module, the features can be efficiently extracted, while improving the generalization ability and accuracy of the model. The latest version of YOLOv8 currently includes five versions: n, s, m, l, and x. Since the YOLOv8s version has fewer parameters and calculations, and the accuracy can meet the needs of small surface target recognition, it is suitable for target detection in telemetry images. Therefore, this embodiment selects YOLOv8s as an improved model.

[0038] In the problem of surface target recognition, illumination changes may cause uneven brightness distribution of the image, thereby affecting the contrast between the structure target and the surrounding background. Ships moving at sea are highly dynamic, moving and changing direction quickly on the water surface. This motion characteristic will cause the edges and contours of the target to become blurred, which may reduce the accuracy of recognition. In addition, the surface environment may present a complex scene, including various background elements such as wind, waves, and flow. These background elements will introduce a lot of redundant information and interference in the image, making it difficult to separate the target from the complex background. The traditional attention mechanism needs to perform spatial pooling operations on the input tensor during the analysis process. In this case, information loss problems are prone to occur, and there is no good interaction between the channels and spaces during calculation. In order to solve the problems of missed detection and multiple detection that may be caused by oil spill pollution sources on the sea surface, this embodiment proposes an improved YOLOv8s algorithm. The algorithm integrates the MCA module based on the YOLOv8s model. The module focuses on extracting position information, enhancing the ability to extract effective features, and helps to improve the problems of missed detection and multiple detection caused by object folding and crossing.

[0039] like Figure 3As shown in Figure 1, the MCA module consists of three parallel branches to model attention in multiple dimensions simultaneously. In the squeeze transformation, the MCA module uses global average and standard deviation pooling to aggregate feature responses across dimensions, and develops a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors. In the excitation transformation, the MCA module captures local feature interactions in a lightweight way instead of using nonlinear strategies for dimensionality reduction to better overcome the trade-off between performance and computational overhead. This design enables the MCA module to achieve multi-dimensional collaborative attention with almost no additional overhead. The top branch of the overall architecture of the MCA module is used to capture the interactions between features in the spatial dimension W. Similarly, the middle branch is used to capture the interactions between features in the spatial dimension H. The bottom branch is responsible for capturing the interactions between channels C. In the first two branches (i.e., the top branch and the middle branch), permutation operations are used to capture the long-range dependencies between the channel dimension and any spatial dimension. Finally, the outputs of the three branches are aggregated by calculating the average in the integration stage.

[0040] In this embodiment, the multi-scale convolution layer is set to three one-dimensional convolution kernels of different sizes, and the convolution kernel sizes are set to d1 = 1 × 1, d2 = 3 × 3, and d3 = 5 × 5 respectively for extracting the short, medium, and long-term features of the input feature sequence. In this process, the activation value at time t in the i-th filter of the first convolution layer is calculated according to the following formula:

[0041]

[0042] Where d represents the convolution kernel size; f represents the activation function; x t,d-t (l-1) Represents the input signal; W i,t (l) ,p i,t (l) ,q i (l) They represent the convolution kernel, convolution output signal and bias coefficient respectively.

[0043] After the input image is processed by the convolutional neural network, feature fusion is performed as required, and then normalized in the following way:

[0044]

[0045] Among them, y (l) represents the activation matrix value, p (l) represents the convolution output signal; E[p (l) ] indicates p (l) The average value of represents the standard deviation; γ (l) , β (l)Both represent adjustment coefficients.

[0046] After determining the activation matrix based on the above analysis, the last layer in the matrix is ​​set as the activation layer, and ReLU is selected as the activation function. After convolution processing, the final output is as follows:

[0047] Y=f(g(c(p s ,p m ,p1)), (3)

[0048] Among them, p s ,p m ,p l They represent the short, medium and long feature sequences obtained after extraction respectively; c(·), g(·) and f(·) represent the connection operation, standardization operation and activation function respectively; Y represents the feature vector.

[0049] Afterwards, the monitoring data collected in S1 is used to train the constructed model to obtain the pollution source model.

[0050] S3. Use the pollution source model to identify the source of marine oil spill pollution.

[0051] After completing the pollution source model through the above steps, the aerial images taken by the new drone are input into the model for identification to determine the location of the marine oil spill pollution source and the specific type of the pollution source.

[0052] Embodiment 2

[0053] The advantages and features of the present invention will be described in detail below in conjunction with the present embodiment.

[0054] In order to enhance the generalization ability and robustness of the model. After setting up the experimental environment, the YOLOv8s network is selected as the main model, and the official recommended parameters are adopted. On this basis, the parameters are appropriately changed. The number of iterations of the training parameters is set to 100, the batch size is set to 8, the random number seed is set to 40, the confidence is set to 0.3, and the intersection and union ratio is set to 0.5. After annotating the collected drone remote sensing data set, it is input into the YOLOv8s algorithm model for training and testing to verify the advantages and disadvantages of the algorithm. In order to verify the impact of the added MCA module on the performance of the YOLOv8s model, this embodiment combines the original model with different attention mechanisms for comparison, and adds the mainstream SE, ECA, and CBAM three attention mechanisms to compare performance indicators. The experimental results are shown in Table 1:

[0055] Table 1

[0056]

[0057] In order to reflect the advantages and balance of the model proposed in this embodiment in terms of detection accuracy and detection speed, a comparative analysis is conducted with the currently popular target detection algorithms. Faster R-CNN, YOLOv4, YOLOv5, YOLOv7, and YOLOv8 are selected for comparison with the improved algorithm of this embodiment, and the comparison is conducted under the same data set and software and hardware conditions to verify the effectiveness of the improvements made to the algorithm model of this embodiment. The analysis experimental results are shown in Table 2:

[0058] Table 2

[0059]

[0060] It can be seen that the improved model has a 9.86% increase in mAP compared to the original algorithm, and the real-time detection also meets the requirements. From the experimental results, it can be seen that although the YOLOv3 model has the fastest detection speed and the least amount of calculation, the detection accuracy is too low. Although the two-stage algorithm FasterR-CNN has a higher detection accuracy, the detection speed is too slow and does not meet the real-time requirements. Although the improved algorithm proposed in this embodiment is slightly higher than the original YOLOv8 algorithm in terms of calculation amount and the detection speed is slightly reduced, it has obvious advantages in detection accuracy, precision and recall rate indicators. The detection accuracy reaches 81.4%, which is 6.7 percentage points higher than the original YOLOv8 algorithm. Considering the computational complexity, detection accuracy and detection speed of each model, the improved model proposed in this embodiment has better performance than the mainstream deep learning detection algorithm.

[0061] Embodiment 3

[0062] This embodiment also provides an oil spill source identification system based on improved YOLOv8s, including: an acquisition module, a construction module and an identification module; the acquisition module is used to collect monitoring data of marine pollution sources, and the marine pollution sources include oil spill pollution sources; the types of monitoring data include drone aerial images; the construction module is used to construct a pollution source identification model based on the monitoring data, and the pollution source model includes an improved YOLOv8s network model, and the improved YOLOv8s network model is a fusion of the MCA module on the basis of the YOLOv8s network model; the identification module is used to use the pollution source model to complete the identification of marine oil spill pollution sources.

[0063] The following will describe in detail how the present invention solves the technical problems in practical work in conjunction with this embodiment.

[0064] First, the drone remote sensing data collected by the acquisition module comes from open source data sets and aerial photography experiment collection. By collecting, organizing and annotating 7 categories of data sets including offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles, the images in these data have various rotation angles and lighting conditions, which helps to train a more robust detection model. When conducting experiments, the 8025 collected data sets were divided into 6200 training sets, 1130 validation sets and 695 test sets. This allocation strategy ensures that the model can learn on rich data and verify accuracy and generalization ability on independent data. In order to minimize the distortion of the image without affecting the detection accuracy, all images are resized to 640×640 and the original aspect ratio is maintained.

[0065] Then, the construction module constructs a pollution source identification model based on the monitoring data. The pollution source model includes an improved YOLOv8s network model. The improved YOLOv8s network model is a YOLOv8s network model that integrates the MCA module.

[0066] Although the recognition algorithm based on convolutional neural network has achieved great success in conventional target detection tasks, the detection effect of surface targets in UAV remote sensing scenarios is not good. The main reason is that in UAV aerial images, the size of objects varies greatly, and the single-layer feature map representation ability of convolutional neural networks is limited. In addition, in aerial images, the background is complex and changeable, and the target object is easily disturbed, which affects the recognition effect. Since the YOLO algorithm has the characteristics of strong versatility, fast, simple, and can adapt to complex sea surface background environments, this embodiment proposes a method for identifying oil spill sources based on improved YOLOv8s for UAV remote sensing image data. The typical representative of the single-stage detection model is YOLOv8. According to practical application experience, the real-time performance of the model is high and the detection accuracy has also reached a high level.

[0067] Among visible light target detection algorithms, the YOLO series of models have high real-time performance and accuracy, and therefore have received widespread attention in the field of target detection. Figure 1As shown in the figure, the YOLOv8 structure can be divided into the backbone feature extraction network, feature enhancement, detection head and other units. The preprocessing operation of the local network module CSP in this network is appropriately simplified, mainly converted into two convolutions, and there is a concept of multi-stack. At the same time, the C3 structure is replaced and converted into a C2f structure, which can better meet the performance requirements of actual applications. The number of channels is appropriately adjusted based on different scales, which has an important role in promoting network learning and also improves the robustness of the algorithm. The CSP module performs two convolutions during operation and multi-stacks different modules to improve the performance of the network. In actual application, the number of channels of the initial convolution is doubled, and then the result is split in half, which is of great significance to improving the calculation speed of the model. The detection head of YOLOv8 uses two independent branches for object classification and predicted bounding box regression. The predicted bounding box regression adopts the DFL and CIoU structures, which improves the detection accuracy and accelerates the convergence of the model.

[0068] YOLOv8 is one of the basic models of the YOLO series. Compared with the original anchor-based detection method, the anchor free method it adopts has higher detection accuracy and detection speed. However, in a complex water surface environment, the floating features of the surface target are complex and the background is diverse. The use of the YOLOv8 detection algorithm may have disadvantages such as positioning error and insufficient target perception. Therefore, in order to solve the above problems, this embodiment proposes an improved YOLOv8 algorithm. Figure 2 As shown, after the feature fusion of the YOLO model, a multi-scale convolution attention module (Multidimensional collaborativeattention, MCA) is set. After the introduction of this module, the features can be extracted efficiently, while improving the generalization ability and accuracy of the model. Since the YOLOv8s version has fewer parameters and calculations, and the accuracy can meet the needs of small surface target recognition, it is suitable for target detection in telemetry images. Therefore, this embodiment selects YOLOv8s as the improved model.

[0069] In the problem of surface target recognition, illumination changes may cause uneven brightness distribution of the image, thereby affecting the contrast between the structure target and the surrounding background. Ships moving at sea are highly dynamic, moving and changing direction quickly on the water surface. This motion characteristic will cause the edges and contours of the target to become blurred, which may reduce the accuracy of recognition. In addition, the surface environment may present a complex scene, including various background elements such as wind, waves, and flow. These background elements will introduce a lot of redundant information and interference in the image, making it difficult to separate the target from the complex background. The traditional attention mechanism needs to perform spatial pooling operations on the input tensor during the analysis process. In this case, information loss problems are prone to occur, and there is no good interaction between the channels and spaces during calculation. In order to solve the problems of missed detection and multiple detection that may be caused by oil spill pollution sources on the sea surface, this embodiment proposes an improved YOLOv8s algorithm. The algorithm integrates the MCA module based on the YOLOv8s model. The module focuses on extracting position information, enhancing the ability to extract effective features, and helps to improve the problems of missed detection and multiple detection caused by object folding and crossing.

[0070] like Figure 3 As shown in Figure 1, the MCA module consists of three parallel branches to model attention in multiple dimensions simultaneously. In the squeeze transformation, the MCA module uses global average and standard deviation pooling to aggregate feature responses across dimensions, and develops a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors. In the excitation transformation, the MCA module captures local feature interactions in a lightweight way instead of using nonlinear strategies for dimensionality reduction to better overcome the trade-off between performance and computational overhead. This design enables the MCA module to achieve multi-dimensional collaborative attention with almost no additional overhead. The top branch of the overall architecture of the MCA module is used to capture the interactions between features in the spatial dimension W. Similarly, the middle branch is used to capture the interactions between features in the spatial dimension H. The bottom branch is responsible for capturing the interactions between channels C. In the first two branches (i.e., the top branch and the middle branch), permutation operations are used to capture the long-range dependencies between the channel dimension and any spatial dimension. Finally, the outputs of the three branches are aggregated by calculating the average in the integration stage.

[0071] In this embodiment, the multi-scale convolution layer is set to three one-dimensional convolution kernels of different sizes, and the convolution kernel sizes are set to d1 = 1 × 1, d2 = 3 × 3, and d3 = 5 × 5 respectively for extracting the short, medium, and long-term features of the input feature sequence. In this process, the activation value at time t in the i-th filter of the first convolution layer is calculated according to the following formula:

[0072]

[0073] Where d represents the convolution kernel size; f represents the activation function; x t,d-t (l-1) Represents the input signal; W i,t (l) ,p i,t (l) ,q i (l) They represent the convolution kernel, convolution output signal and bias coefficient respectively.

[0074] After the input image is processed by the convolutional neural network, feature fusion is performed as required, and then normalized in the following way:

[0075]

[0076] Among them, y (l) represents the activation matrix value, p (l) represents the convolution output signal; E[p (l) ] indicates p (l) The average value of represents the standard deviation; γ (l) , β (l) Represents the adjustment coefficient.

[0077] After determining the activation matrix based on the above analysis, the last layer in the matrix is ​​set as the activation layer, and ReLU is selected as the activation function. After convolution processing, the final output is as follows:

[0078] Y=f(g(c(p s ,p m ,p1)), (6)

[0079] Among them, p s ,p m ,p l They represent the short, medium and long feature sequences obtained after extraction respectively; c(·), g(·) and f(·) represent the connection operation, standardization operation and activation function respectively; Y represents the feature vector.

[0080] Afterwards, the monitoring data collected by the acquisition module is used to train the constructed model to obtain the pollution source model.

[0081] Finally, the identification module uses the pollution source model to complete the identification of the source of marine oil spill pollution.

[0082] After completing the pollution source model through the above steps, the aerial images taken by the new drone are input into the model for identification to determine the location of the marine oil spill pollution source and the specific type of the pollution source.

[0083] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. The oil spill source identification method based on improved YOLOv8s is characterized by the following steps: include: Collecting monitoring data of marine pollution sources, wherein the marine pollution sources include oil spill pollution sources; the types of monitoring data include drone aerial images; Based on the monitoring data, a pollution source identification model is constructed, wherein the pollution source model includes an improved YOLOv8s network model, wherein the improved YOLOv8s network model is a YOLOv8s network model integrated with an MCA module; the introduced MCA module is composed of three parallel branches, wherein the top branch is used to capture the interaction between features in the spatial dimension W; the middle branch is used to capture the interaction between features in the spatial dimension H; and the bottom branch is responsible for capturing the interaction between channels C; In the top branch and the middle branch, a permutation operation is used to capture the long-range dependency between the channel dimension and any spatial dimension; finally, the outputs of the three branches are aggregated by calculating the average in the integration stage; The pollution source model is used to identify the source of marine oil spill pollution.

2. The oil spill source identification method based on improved YOLOv8s according to claim 1 is characterized in that: The collected monitoring data are sorted and labeled, and the monitoring data include offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles; the collected monitoring data are divided into a training set, a validation set and a test set according to a preset ratio.

3. The oil spill source identification method based on improved YOLOv8s according to claim 1, characterized in that: The MCA module adopts global average and standard deviation pooling to aggregate cross-dimensional feature responses in the squeeze transformation, and uses a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors; The MCA module captures local feature interactions in a lightweight manner during excitation transformation, overcoming the trade-off between performance and computational overhead.

4. An oil spill source identification system based on improved YOLOv8s, the system is used to implement the method described in any one of claims 1 to 3, characterized in that: include: Acquisition module, construction module and recognition module; The acquisition module is used to collect monitoring data of marine pollution sources, wherein the marine pollution sources include oil spill pollution sources; the types of monitoring data include drone aerial images; The construction module is used to construct a pollution source identification model based on the monitoring data, and the pollution source model includes an improved YOLOv8s network model, and the improved YOLOv8s network model is a fusion of the MCA module on the basis of the YOLOv8s network model; the introduced MCA module is composed of three parallel branches, wherein the top branch is used to capture the interaction between features in the spatial dimension W; the middle branch is used to capture the interaction between features in the spatial dimension H; the bottom branch is responsible for capturing the interaction between channels C; in the top branch and the middle branch, a permutation operation is used to capture the long-range dependency between the channel dimension and any spatial dimension; finally, in the integration stage, the outputs of the three branches are aggregated by calculating the average; The identification module is used to use the pollution source model to complete the identification of the marine oil spill pollution source.

5. The oil spill source identification system based on improved YOLOv8s according to claim 4 is characterized in that: The workflow of the acquisition module includes: organizing and labeling the collected monitoring data, which includes offshore oil drilling platforms, buoys, fishing boats, container ships, passenger ships, warships and other surface obstacles; dividing the collected monitoring data into training sets, verification sets and test sets according to preset proportions.

6. The oil spill source identification system based on improved YOLOv8s according to claim 4, characterized in that: The MCA module adopts global average and standard deviation pooling to aggregate cross-dimensional feature responses in the squeeze transformation, and uses a combination mechanism to adaptively merge the features of average pooling and standard deviation pooling to enhance the representation of feature descriptors; The MCA module captures local feature interactions in a lightweight manner during excitation transformation, overcoming the trade-off between performance and computational overhead.

Images