Multi-platform sea floating garbage monitoring system and method
Through a multi-platform sea drift garbage monitoring system, combined with shore-based and air-based monitoring modules, the YOLOv5 model is used to identify sea drift garbage, which solves the problem of low efficiency of traditional monitoring methods and realizes large-scale and all-weather sea drift garbage monitoring and classification , reducing monitoring costs and manpower dependence.
Patent Information
- Application Number
- CN202411316679.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2025-01-03
AI Technical Summary
Traditional sea drift garbage monitoring methods are inefficient, cannot achieve large-scale, all-weather, and multi-scale real-time monitoring. Due to weather and sight limitations, it is difficult to effectively identify and deal with garbage on the coastline and offshore sea surfaces.
Adopt a multi-platform sea drift waste monitoring system, combined with shore-based and air-based monitoring modules, and use cameras to collect image data, and real-time identification and positioning through data processing platforms and target detection models, including shore-based monitoring modules, air-based monitoring modules and data The processing platform uses the YOLOv5 model for garbage identification and classification, and conducts sea area monitoring through drones.
Large-scale and all-weather sea drift garbage monitoring has been achieved, which has reduced monitoring labor costs, improved monitoring efficiency, and can identify and classify sea drift garbage in real time in harsh climates, reducing dependence on labor.
Smart Images
Figure CN119251710A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of optical monitoring and image processing, and in particular to a multi-platform marine debris monitoring system and method. Background Art
[0002] In recent years, with the vigorous development of human marine activities, marine debris has become a problem that cannot be ignored. According to statistics, 8 million tons of garbage enter the ocean every year, of which 80% is land-based and 20% is marine-based. A large amount of marine debris distributed on the coastline and offshore not only affects the marine landscape, but also produces a large amount of toxic and harmful substances, causing inestimable harm to aquaculture and fisheries, and hindering the economic development of coastal areas. Therefore, it is necessary to conduct large-scale, real-time monitoring of garbage on the coastline and offshore to protect the environment and promote economic development.
[0003] Traditional monitoring of floating garbage at sea mainly relies on regular manual patrols along the coastline, reporting the garbage after it is found, and the handling personnel salvaging the garbage after arriving at the scene. However, due to the long coastline and complex geographical conditions, manual patrols are inefficient and costly; and manual patrols can only be observed from the ground, making it difficult to find garbage in blind spots; in addition, it is seriously affected by the weather. There is no patrol under bad weather conditions and at night, and it can only wait for conditions to improve before handling it, and it is impossible to achieve all-weather monitoring; in addition, floating garbage at sea is a random and sporadic event, making it difficult to monitor and deal with it in real time. In summary, conventional methods of monitoring floating garbage at sea have major defects, making it difficult to achieve large-scale, all-weather, and multi-scale real-time monitoring. Summary of the invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a multi-platform marine debris monitoring system, which can realize large-scale, all-weather real-time monitoring of the coast and offshore waters, identification and early warning of marine debris, provide intelligent technical means for the management of marine debris, and reduce the labor cost of monitoring.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] A multi-platform marine debris monitoring system includes a shore-based monitoring module, an air-based monitoring module and a data processing platform. The shore-based monitoring module includes multiple cameras and a shore-based control computer. The cameras collect image data or video data of the monitored sea area, and then transmit the collected data to the shore-based control computer, which then transmits the collected data to the data processing platform. The air-based monitoring module includes an unmanned aerial vehicle and an air-based control computer. The unmanned aerial vehicle collects image data or video data of the monitored sea area, and then transmits the collected data to the air-based control computer, which then transmits the collected data to the data processing platform. The data processing platform processes the returned data through a trained target detection model, identifies and locates marine debris, and then notifies the user of the identification results.
[0007] Preferably, the training process of the target detection model is as follows: collect garbage images in various scenarios, filter out images containing marine debris and construct a sample data set; divide the sample data set into a training set, a test set and a validation set, and select and mark the marine debris on the images in the sample data set to generate labels; the target detection model is constructed based on YOLOv5, and the network is optimized; during the model training process, the model is trained using the training set, and the validation set is used for validation after each training round to detect the performance of the model on unseen data to prevent overfitting; the model is evaluated, and the trained model is finally evaluated using the test set, including the accuracy, recall rate, F1 score and other relevant performance indicators of the model. The model is deployed and applied, and after evaluation, the optimal model is determined and deployed in the data processing platform for real-time detection and classification of marine debris.
[0008] Preferably, the network optimization of the target detection model includes: Pooling layer optimization strategy: Introduce the Softpool pooling method, which calculates the weighted average of the elements in each pooling window, and its weight is the exponential value of these elements, with the hyperparameter α as the base, so as to flexibly balance the maximum pooling and average pooling in the convolutional neural network; by adjusting α, Softpool can optimize between performance and sensitivity, making feature extraction more efficient and smooth. Data enhancement: Perform data enhancement on the training data, including random cropping, rotation, scaling and color transformation, so as to increase the diversity of the training data. Transfer learning: In order to improve the efficiency of model training and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale data set, accelerating the training process of the target detection model.
[0009] More preferably, the shore-based control computer and the air-based control computer perform segmentation processing on the collected image data, divide the image data into N equal parts, and number each segmented small image.
[0010] Based on the same inventive concept, the present invention also provides a multi-platform marine debris monitoring method.
[0011] A multi-platform marine floating garbage monitoring method is provided, which deploys a shore-based monitoring module, an air-based monitoring module and a data processing platform, wherein the shore-based monitoring module comprises a plurality of cameras and a shore-based control computer, and the air-based monitoring module comprises an unmanned aerial vehicle and an air-based control computer, and then performs the following steps: the camera collects image data or video data of the monitored sea area, and then transmits the collected data to the shore-based control computer, and then transmits the collected data to the data processing platform through the shore-based control computer; the unmanned aerial vehicle collects image data or video data of the monitored sea area, and then transmits the collected data to the air-based control computer, and then transmits the collected data to the data processing platform through the air-based control computer; the data processing platform processes the returned data through a trained target detection model, identifies and locates the marine floating garbage, and then notifies the user of the identification result.
[0012] Preferably, the training process of the target detection model is as follows: collect garbage images in various scenarios, filter out images containing marine debris and construct a sample data set; divide the sample data set into a training set, a test set and a validation set, and select and mark the marine debris on the images in the sample data set to generate labels; the target detection model is constructed based on YOLOv5, and the network is optimized; during the model training process, the model is trained using the training set, and the validation set is used for validation after each training round to detect the performance of the model on unseen data to prevent overfitting; the model is evaluated, and the trained model is finally evaluated using the test set, including the accuracy, recall rate, F1 score and other relevant performance indicators of the model. The model is deployed and applied, and after evaluation, the optimal model is determined and deployed in the data processing platform for real-time detection and classification of marine debris.
[0013] Preferably, the network optimization of the target detection model includes: Pooling layer optimization strategy: Introduce the Softpool pooling method, which calculates the weighted average of the elements in each pooling window, and its weight is the exponential value of these elements, with the hyperparameter α as the base, so as to flexibly balance the maximum pooling and average pooling in the convolutional neural network; by adjusting α, Softpool can optimize between performance and sensitivity, making feature extraction more efficient and smooth. Data enhancement: Perform data enhancement on the training data, including random cropping, rotation, scaling and color transformation, so as to increase the diversity of the training data. Transfer learning: In order to improve the efficiency of model training and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale data set, accelerating the training process of the target detection model.
[0014] More preferably, the shore-based control computer and the air-based control computer perform segmentation processing on the collected image data, divide the image data into N equal parts, and number each segmented small image.
[0015] The present invention has the following beneficial effects:
[0016] 1. The monitoring system and method of the present invention use artificial intelligence algorithms to identify marine debris, which can effectively monitor the occasional event of marine debris and classify common garbage such as foam, driftwood, plastic bottles, plastic bags, etc.
[0017] 2. The monitoring system and method of the present invention use artificial intelligence algorithms to identify marine debris, can achieve large-scale all-weather monitoring, are not affected by meteorological conditions, can effectively improve monitoring efficiency, and reduce dependence on manual labor.
[0018] 3. The monitoring system and method of the present invention are not limited by the data source, can process data from different sources, and realize real-time monitoring on multiple platforms. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a schematic diagram of the monitoring system principle of the present invention;
[0020] Figure 2 It is a schematic diagram of the monitoring system of the present invention;
[0021] Figure 3 This is a schematic diagram of the shore-based camera deployment and drone route design of the present invention;
[0022] Figure 4 and Figure 5 A physical picture collected by the drone of the present invention;
[0023] Figure 6 The physical image collected by the shore-based camera of the present invention;
[0024] Figure 7 It is a schematic diagram of the preprocessing of the collected images of the present invention. DETAILED DESCRIPTION
[0025] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0026] Embodiment 1
[0027] See also Figure 1 and Figure 2 A multi-platform marine floating garbage monitoring system includes a shore-based monitoring module 10, an air-based monitoring module 20 and a data processing platform 30. The shore-based monitoring module 10 includes a plurality of cameras 11 and a shore-based control computer 12. The cameras 11 are deployed at key locations such as docks, estuaries and coastlines. The cameras 11 can obtain 2560×1440, 24-bit high-definition monitoring image data or video data, and communicate with the back-end control computer through a data line to transmit the acquired data to the shore-based control computer 12. The shore-based control computer 12 contains a user interaction software system. The software system realizes the hardware module control, data transmission and storage of the entire shore-based monitoring module. The shore-based monitoring module can control whether the camera 11 is working, whether to save the monitoring data, and whether to return the on-site monitoring data to the data processing platform 30. The user interaction software system is compiled by Python language, which is easy to expand and can be improved according to the on-site situation. The collected data is transmitted to the data processing platform 30 through the shore-based control computer 12.
[0028] The air-based monitoring module 20 includes an unmanned aerial vehicle 21 and an air-based control computer 22. The unmanned aerial vehicle 21 is equipped with an optical camera 11 and can fly and patrol along the coastline. The camera 11 can obtain 2560×1440, 24-bit high-definition monitoring image data or video data. After the unmanned aerial vehicle 21 returns, it can use a data cable to communicate with the air-based control computer 22 to transmit the acquired data to the air-based control computer 22. The air-based control computer 22 contains a user interaction software system. The software system realizes the transmission and storage of data of the entire air-based monitoring module. The air-based monitoring module can save monitoring data and return on-site monitoring data to the data processing module. The user interaction software system is compiled by Python language, which is easy to expand and can be improved according to the on-site situation. The collected data is transmitted to the data processing platform 30 through the air-based control computer 22.
[0029] The data processing platform 30 is a high-performance workstation, which contains a user interactive software system and can communicate with the shore-based monitoring module 10 and the air-based monitoring module 20 through the 4G communication function to obtain on-site monitoring data. The data processing platform 30 processes the data sent back through the trained target detection model, identifies the possible target objects, and saves the identification results when there is no marine debris. When there is marine debris, the marine debris is identified and located, and the identification results can be viewed in real time, and the identification results are notified to the user and saved at the same time.
[0030] Through the above two modules, the ground and air perspectives are obtained, and the monitoring data is monitored 24 hours a day. In order to reduce the pressure on the data processing platform 30, speed up the processing speed and improve work efficiency. On the control computers of the shore-based monitoring module 10 and the air-based monitoring module 20, the acquired data is pre-processed: the shore-based control computer 12 and the air-based control computer 22 segment the collected image data, divide the image data into N equal parts, and number each segmented small image. Please refer to Figure 7 For example, the obtained image is divided into the upper left corner numbered 1, the upper right corner numbered 2, the lower left corner numbered 3, and the lower right corner numbered 4. See the segmentation diagram Figure 5 The segmented image can effectively reduce the pressure on the data processing platform 30.
[0031] Preferably, the training process of the target detection model is:
[0032] Step 1: Collect garbage images in various scenarios, filter out images containing marine debris, and construct a sample dataset;
[0033] Step 2: Divide the sample data set into a training set, a test set, and a validation set, select and mark the marine debris on the images in the sample data set, and generate labels;
[0034] Step 3: The target detection model is built based on YOLOv5, and the network is optimized.
[0035] Step 3.1 The maximum pooling (Maxpool) used in the original YOLOv5 network can reduce the estimated mean shift caused by the convolution layer parameter error, and can retain more texture features, but it is easy to ignore the detail feature information. Therefore, this embodiment introduces the Softpool pooling method, which uses Softmax weighting and has differentiable characteristics, so that the network can retain the basic properties of the input image information and expand the feature information. The Softpool pooling method calculates the weighted average of the elements in each pooling window, and its weight is the exponential value of these elements, with the hyperparameter α as the base, so as to flexibly balance the maximum pooling and average pooling in the convolutional neural network; by adjusting α, Softpool can optimize between performance and sensitivity, making feature extraction more efficient and smooth.
[0036] Step 3.2 Data enhancement: Perform data enhancement on the training data, including random cropping, rotation, scaling, and color transformation, to increase the diversity of the training data.
[0037] Step 3.3 Transfer learning: In order to improve the efficiency of model training and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale dataset to accelerate the training process of the target detection model. The process is: select a model that has been trained on a large-scale dataset, load its pre-trained weights, and migrate these weights to the new target detection model. Some layers are selected to freeze to retain common features, while continuing to train on the target detection task to learn specific features. The pre-trained model provides a large amount of low-level and mid-level feature information that has been learned, providing a good initialization for the target detection model, thereby accelerating the training process and improving the final performance.
[0038] Step 4: During the model training process, the model is trained using the training set and validated using the validation set after each training round to detect the performance of the model on unseen data and prevent overfitting;
[0039] Step 5: Evaluate the model and use the test set to perform a final evaluation of the trained model, including the accuracy, recall rate, F1 score, and other relevant performance indicators of the model.
[0040] Step 6: Deploy and apply the model. After evaluation, determine the optimal model and deploy it in the data processing platform 30 for real-time detection and classification of marine debris.
[0041] Through the above steps 1 to 6, an efficient marine debris detection model is constructed, which realizes the accurate identification and classification of various marine debris.
[0042] The implementation process of the present invention is now described by taking a 5-kilometer-long coast to be monitored as an example.
[0043] according to Figure 2 , Figure 3 as well as Figure 6 , along the coastline, multiple shore-based surveillance cameras 11 are arranged. The height of the camera 11 is about 2-3 meters, and the tilt angle is 35° to 55°. It is adjusted according to the on-site situation to ensure that the sea surface can be monitored in the largest area. In addition, the same camera 11 is additionally set for the docks, estuaries and other locations with frequent daily activities within the coastline for key monitoring. Each camera 11 is numbered in sequence with three digits and controlled by user interactive software. The control software captures the surveillance camera 11 screen every 5 minutes and names it according to certain naming rules. The naming rules are: the first 6 digits are year, month and day information; the 7th to 12th digits are time information; the 13th digit is the module identification code, the shore-based module is 0, and the air-based module is 1. If there is other source data, it can be expanded in sequence; the 14th to 16th digits are station information. For example, the data captured by the shore-based camera 111 at 8:07:06 on January 1, 2022 can be named 2201010807060001. The data captured by the user interaction software is stored locally for the data processing platform 30 to call.
[0044] See also Figures 2 to 5 , set up drone 21 routes along the coastline, drone 21 flight altitude of 20 meters, speed of 60km / h to 80km / h, ensure coverage of 15 meters on both sides of the coastline, and adjust according to the on-site situation to ensure the maximum area of surveillance of the sea surface. Encode the data obtained by different take-off times and different drones 21, and the encoding rules are as follows: the first 6 digits are year, month, and day information; the 7th to 8th digits are take-off information, one flight per hour, a total of 24 flights per day, and the 9th to 12th digits are serial numbers; the 13th digit is the module identification code, the shore-based module is 0, and the air-based module is 1. If there are other data sources, they can be expanded in sequence; the 14th to 16th digits are the three-digit drone 21 number. For example, the first image obtained by drone 21 numbered 001, which took off at 8 o'clock on January 1, 2023, is numbered 2301010800011001. The data obtained by the user interaction software will be stored locally for the data processing platform 30 to call.
[0045] The data processing platform 30 reads the pre-processed image data through network communication, imports the image data into the target detection model, identifies the possible target objects, and saves the identification results into the task log. At the same time, if a large amount of floating garbage is detected, an alarm will be immediately issued to notify the on-duty personnel to dispose of the garbage immediately.
[0046] The monitoring system of the present invention uses artificial intelligence algorithms to identify floating garbage at sea, can effectively monitor the occasional floating garbage at sea, and classify common garbage such as foam, driftwood, plastic bottles, plastic bags, etc. At the same time, the present invention can achieve large-scale all-weather monitoring, is not affected by meteorological conditions, can effectively improve monitoring efficiency, and reduce dependence on manual labor. The monitoring system of the present invention is not limited by the data source, can process data from different sources, and realize multi-platform real-time monitoring.
[0047] Based on the same inventive concept, the second embodiment provides the method steps executed by the system of the first embodiment.
[0048] Embodiment 2
[0049] A multi-platform marine floating garbage monitoring method is provided, which deploys a shore-based monitoring module 10, an air-based monitoring module 20 and a data processing platform 30, wherein the shore-based monitoring module 10 comprises a plurality of cameras 11 and a shore-based control computer 12, and the air-based monitoring module 20 comprises an unmanned aerial vehicle 21 and an air-based control computer 22, and then performs the following steps: the camera 11 collects image data or video data of the monitored sea area, and then transmits the collected data to the shore-based control computer 12, and then transmits the collected data to the data processing platform 30 through the shore-based control computer 12; the unmanned aerial vehicle 21 collects image data or video data of the monitored sea area, and then transmits the collected data to the air-based control computer 22, and then transmits the collected data to the data processing platform 30 through the air-based control computer 22; the data processing platform 30 processes the returned data through a trained target detection model, identifies and locates the marine floating garbage, and then notifies the user of the identification result.
[0050] Preferably, the training process of the target detection model is as follows: collecting garbage images in various scenes, screening out images containing marine debris and constructing a sample data set; dividing the sample data set into a training set, a test set and a validation set, and selecting and marking the marine debris on the images in the sample data set to generate labels; the target detection model is constructed based on YOLOv5, and the network is optimized; during the model training process, the model is trained using the training set, and the validation set is used for validation after each training round to detect the performance of the model on unseen data to prevent overfitting; the model is evaluated, and the trained model is finally evaluated using the test set, including the accuracy, recall rate, F1 score and other relevant performance indicators of the evaluation model. The model is deployed and applied, and after evaluation, the optimal model is determined and deployed in the data processing platform 30 for real-time detection and classification of marine debris.
[0051] Preferably, the network optimization of the target detection model includes: Pooling layer optimization strategy: Introducing the Softpool pooling method, which uses Softmax weighting and has differentiable features, so that the network can retain the basic properties of the input image information and expand the feature information. Data enhancement: Perform data enhancement on the training data, including random cropping, rotation, scaling and color transformation, so as to increase the diversity of the training data. Transfer learning: In order to improve the efficiency of model training and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale data set, which accelerates the training process of the target detection model.
[0052] More preferably, the shore-based control computer 12 and the air-based control computer 22 perform segmentation processing on the collected image data, divide the image data into N equal parts, and number each segmented small image.
[0053] Since the method described in the second embodiment of the present invention is a method step for implementing the system of the first embodiment of the present invention, those skilled in the art can understand the specific implementation of the system based on the system described in the first embodiment of the present invention, and therefore, no further description is given here. All technical means described in the first embodiment of the present invention belong to the scope of protection of the present invention.
[0054] The above description is only a specific implementation mode of the present invention, and does not limit the patent scope of the present invention. Any equivalent structural transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A multi-platform marine debris monitoring system, characterized by: It includes shore-based monitoring module, air-based monitoring module and data processing platform; The shore-based monitoring module includes a plurality of cameras and a shore-based control computer. The cameras collect image data or video data of the monitored sea area, and then transmit the collected data to the shore-based control computer, and the shore-based control computer transmits the collected data to the data processing platform; The air-based monitoring module includes an unmanned aerial vehicle and an air-based control computer. The unmanned aerial vehicle collects image data or video data of the monitored sea area, and then transmits the collected data to the air-based control computer, and the air-based control computer transmits the collected data to the data processing platform; The data processing platform processes the returned data through a trained target detection model, identifies and locates marine debris, and then notifies the user of the identification results.
2. A multi-platform marine debris monitoring system according to claim 1, characterized in that: The training process of the target detection model is: Collect garbage images in various scenarios, filter out images containing marine debris and build a sample dataset; The sample dataset is divided into a training set, a test set, and a validation set. The marine debris on the images in the sample dataset is selected and categorized to generate labels. The target detection model is built based on YOLOv5, and the network is optimized; During the model training process, the model is trained using the training set, and the validation set is used for validation after each training round to detect the performance of the model on unseen data and prevent overfitting; Evaluate the model and use the test set to perform a final evaluation of the trained model, including the accuracy, recall rate, F1 score, and other relevant performance indicators of the model; The model is deployed and applied, and after evaluation, the optimal model is determined and deployed in the data processing platform for real-time detection and classification of marine debris.
3. A multi-platform marine debris monitoring system according to claim 2, characterized in that: The network optimization of the target detection model includes: Pooling layer optimization strategy: Introduce the Softpool pooling method, which calculates the weighted average of the elements in each pooling window. The weight is the exponential value of these elements, with the hyperparameter α as the base, so as to flexibly balance the maximum pooling and average pooling in the convolutional neural network; by adjusting α, Softpool can optimize between performance and sensitivity, making feature extraction more efficient and smooth. Data augmentation: Perform data augmentation on training data, including random cropping, rotation, scaling, and color transformation, to increase the diversity of training data; Transfer learning: In order to improve the training efficiency of the target detection model and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale data set, which accelerates the training process of the target detection model.
4. The multi-platform marine debris monitoring system according to claim 1, characterized in that: The shore-based control computer and the air-based control computer perform segmentation processing on the collected image data, divide the image data into N equal parts, and number each segmented small image.
5. A multi-platform marine debris monitoring method, characterized in that: Deploy a shore-based monitoring module, an air-based monitoring module and a data processing platform, wherein the shore-based monitoring module includes multiple cameras and a shore-based control computer, and the air-based monitoring module includes a drone and an air-based control computer, and then perform the following steps: The camera collects image data or video data of the monitored sea area, and then transmits the collected data to the shore-based control computer, and the shore-based control computer transmits the collected data to the data processing platform; The drone collects image data or video data of the monitored sea area, and then transmits the collected data to the air-based control computer, which transmits the collected data to the data processing platform through the air-based control computer; The data processing platform processes the returned data through a trained target detection model, identifies and locates marine debris, and then notifies the user of the identification results.
6. A multi-platform marine debris monitoring method according to claim 5, characterized in that: The training process of the target detection model is: Collect garbage images in various scenarios, filter out images containing marine debris and build a sample dataset; The sample dataset is divided into a training set, a test set, and a validation set. The marine debris on the images in the sample dataset is selected and categorized to generate labels. The target detection model is built based on YOLOv5, and the network is optimized; During the model training process, the model is trained using the training set, and the validation set is used for validation after each training round to detect the performance of the model on unseen data and prevent overfitting; Evaluate the model and use the test set to perform a final evaluation of the trained model, including the accuracy, recall rate, F1 score, and other relevant performance indicators of the model; The model is deployed and applied, and after evaluation, the optimal model is determined and deployed in the data processing platform for real-time detection and classification of marine debris.
7. A multi-platform marine debris monitoring method according to claim 6, characterized in that: The network optimization of the target detection model includes: Pooling layer optimization strategy: Introduce the Softpool pooling method, which calculates the weighted average of the elements in each pooling window, with the weight being the exponential value of these elements, based on the hyperparameter α, so as to flexibly balance the maximum pooling and average pooling in the convolutional neural network; by adjusting α, Softpool can optimize between performance and sensitivity, making feature extraction more efficient and smooth; Data augmentation: Perform data augmentation on training data, including random cropping, rotation, scaling, and color transformation, to increase the diversity of training data; Transfer learning: In order to improve the efficiency of model training and the performance of the final model, a pre-trained model is used for initialization. The pre-trained model has learned rich feature information on a large-scale data set, which accelerates the training process of the target detection model.
8. A multi-platform marine debris monitoring method according to claim 5, characterized in that: The shore-based control computer and the air-based control computer perform segmentation processing on the collected image data, divide the image data into N equal parts, and number each segmented small image.