A multi-scale ocean floating garbage detection method resistant to wave interference

Abstract

The application belongs to the technical field of computer vision and marine environment monitoring, and discloses a multi-scale marine floating garbage detection method resisting wave interference. The method is based on YOLOv11 as a basic framework, and is optimized from three key dimensions of loss function, feature fusion and detection head. A WIoU v3 loss function based on a dynamic non-monotonic focusing mechanism is adopted to adaptively adjust the sample gradient gain and suppress the interference of low-quality samples. A pixel-level dynamic weight remapping is realized through an ASFF adaptive feature fusion module to eliminate multi-scale feature conflicts and filter wave noise. A task-aware dual-path decoupling detection head is designed in combination with a double-flow context enhancement unit to block the gradient interference of classification and regression tasks and balance local details and global context information. The application can accurately identify marine floating garbage under complex sea conditions such as strong light reflection and wave breaking, and can be deployed on mobile terminals such as unmanned ships and unmanned aerial vehicles to realize real-time and efficient monitoring.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and marine environmental monitoring technology, and in particular to a multi-scale marine floating debris detection method that is resistant to wave interference. Background Technology

[0002] Marine debris pollution has become a global ecological problem, posing a serious threat to marine ecosystems, shipping safety, and the coastal environment. Timely and accurate monitoring of the distribution and types of marine debris is a prerequisite for pollution control and prevention efforts. With the development of unmanned surface vessels and drones, machine vision-based mobile marine monitoring has become the mainstream solution, but it still faces many challenges in practical applications.

[0003] Existing methods for detecting marine floating debris are mostly based on traditional target detection algorithms, which have significant limitations under complex sea conditions. On the one hand, interference factors such as white foam generated by breaking waves, target blurring due to strong light reflection, and feature weakening caused by water turbidity can easily lead to false alarms or missed detections. On the other hand, marine floating debris has multi-scale and irregular morphological characteristics, and some tiny debris has high similarity to the background texture, making it difficult for traditional algorithms to effectively balance the contradiction between multi-scale feature extraction and noise suppression. In addition, the loss functions of existing algorithms often adopt monotonic focusing strategies, which are prone to overfitting low-quality samples and generating harmful gradients. The coupled architecture of the detection head's classification and regression tasks can lead to gradient interference, further reducing detection accuracy and generalization ability.

[0004] To address the aforementioned issues, this invention proposes a highly efficient method for detecting floating marine debris that possesses strong resistance to wave interference and can adapt to targets at multiple scales, thereby meeting the practical needs of real-time monitoring by mobile terminals. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies in detecting marine floating debris under complex sea conditions, and provides a multi-scale marine floating debris detection method that is resistant to wave interference. This method enables accurate identification of multi-scale, irregular floating debris, improves the robustness and real-time detection performance of the model under complex environments such as wave interference and changes in lighting, and is suitable for deployment requirements of mobile terminals such as unmanned ships and drones.

[0006] A multi-scale marine floating debris detection method resistant to wave interference includes the following steps:

[0007] S1. Constructing an interference-resistant marine floating debris dataset: Raw sea surface images under different sea conditions were collected by unmanned vessels and drones. After manual cleaning and CVAT annotation, a dataset containing seven types of floating debris was obtained. A triple strategy of Mosaic stitching, HSV illumination transformation, and Mixup hybrid enhancement was adopted to expand the dataset and improve the noise resistance and robustness of the model.

[0008] S2. Construct the MFL-YOLO optimization model: Based on YOLOv11, integrate three major modules:

[0009] (1) The WIoU v3 loss function is adopted, and the sample gradient gain is adaptively adjusted based on distance attention and dynamic non-monotonic focusing mechanism;

[0010] (2) An ASFF adaptive feature fusion module is embedded in the neck to achieve pixel-level dynamic weight remapping, suppress wave noise and eliminate multi-scale feature conflicts;

[0011] (3) Design a task-aware dual-path decoupled detection head, combined with a dual-stream context enhancement unit, to block gradient interference in classification and regression tasks;

[0012] S3. Model Training: Train the MFL-YOLO model based on the constructed dataset, iteratively optimize the parameters using WIoU v3 as the loss function, and save the optimal model weights;

[0013] S4. Mobile Terminal Deployment and Real-time Inference: After converting the trained model format, deploy it to the mobile terminal of the unmanned vessel / drone. Collect sea surface images in real time and input them into the model. Remove redundant detection boxes through non-maximum suppression algorithm, and output the category and location coordinates of floating debris to achieve real-time monitoring against wave interference.

[0014] Furthermore, the calculation process of the WIoU v3 loss function includes:

[0015] a. Constructing a base loss function based on distance attention :

[0016]

[0017]

[0018] In the formula, =1-IoU,(x,y),( ) represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively, W g H g Let the width and height be the minimum bounding rectangle;

[0019] b. Define outlier β to measure sample quality:

[0020]

[0021] In the formula, The regression loss for the current sample; The average IoU loss for the current batch, and The calculation process was interrupted. The gradient is backpropagated so that it is used only as a reference coefficient for dynamic weights;

[0022] Construct the non-monotonic gradient gain coefficient r to obtain the final loss function:

[0023]

[0024]

[0025] In the formula, α and δ are hyperparameters. For the final loss function

[0026] Furthermore, the implementation process of the ASFF adaptive feature fusion module includes:

[0027] a. Unify the features of the three layers of shallow layer P3, middle layer P4 and deep layer P5 to the same spatial resolution by upsampling or downsampling, and calculate the weight coefficients α, β and γ of the three layers P3, P4 and P5, and α+β+γ=1. , , These represent feature vectors from the three levels P3, P4, and P5, respectively, adjusted to the resolution of the current level l.

[0028] b. Learn pixel-level weight coefficients by normalizing using the Softmax function. , , :

[0029]

[0030] In the formula, , , These are the control parameters learned through backpropagation of the network; similarly, we can obtain... and ;

[0031] c. Fusion features The calculation is a weighted sum of features from each layer:

[0032]

[0033] The weights for feature fusion are dynamically generated for each pixel in the image, adaptively suppressing background noise caused by near-field wave reflections.

[0034] Furthermore, the implementation process of the task-aware dual-path decoupling detection head includes:

[0035] a. Divide the input feature stream into independent classification and regression branches. Extract local and global features in parallel within each branch. The local branch uses a 3×3 standard convolution function to extract edge detail features. The global branch uses a dilated convolution function with a dilation rate of d=2 to extract environmental context features. The mathematical expression is as follows:

[0036]

[0037]

[0038] b. Concatenate the two feature streams along the channel dimension and apply a weight matrix. Convolutional layers are used for dimensionality reduction and feature interaction to generate fused features. :

[0039]

[0040] In the formula, This represents a channel splicing operation. The weight parameters represent the weights of the 1×1 convolution kernel;

[0041] c. Activation function via Sigmoid Construct a self-attention gating mechanism and fuse features After being weighted by its own Sigmoid activation value, it is compared with the original input. By performing residual superposition, the final output is obtained:

[0042]

[0043] In the formula, A task-aware mechanism is constructed, which automatically suppresses background noise responses in the fused features that are irrelevant to the current task by using nonlinear activation.

[0044] Furthermore, the seven types of floating waste include plastic bottles, plastic bags, foam, fishing nets, glass bottles, aluminum cans, and other fragments. The dataset is divided into training set:validation set:test set = 7:2:1, and the sample size of the training set is expanded to 4 times the original size after enhancement.

[0045] Furthermore, the model training parameters were set as follows: batch size 16, initial learning rate 0.001, cosine annealing learning rate scheduling, weight decay coefficient 0.0005, training epochs 150, confidence threshold for nonmaximum suppression algorithm 0.5, and IOU threshold 0.45.

[0046] Furthermore, the method is deployed on unmanned ships and drone mobile terminals.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. Strong anti-interference capability: Through the ASFF anti-noise fusion module and triple data enhancement strategy, it effectively suppresses background interference such as wave foam and light reflection, significantly reduces the false alarm rate, and adapts to complex dynamic sea conditions.

[0049] 2. High detection accuracy: The WIoU v3 loss function solves the problem of gradient interference from low-quality samples, and the task-aware dual-path decoupled detection head balances the requirements of classification and regression tasks, significantly improving the recognition accuracy and positioning accuracy of multi-scale, irregular floating garbage.

[0050] 3. Excellent real-time performance: The model is based on the YOLOv11 lightweight architecture and is accelerated by quantization so that it can be deployed on mobile terminals. The inference speed meets the requirements of real-time monitoring and is suitable for routine inspections of unmanned ships and drones.

[0051] 4. Good generalization ability: The diverse datasets and adaptive optimization mechanism enable the model to adapt to the detection needs of different sea conditions and different types of floating debris, making it highly practical and providing technical support for marine environmental monitoring. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a diagram of the overall network architecture of the MFL-YOLO of this invention;

[0054] Figure 2 This is a technical roadmap of the present invention;

[0055] Figure 3 A statistical distribution diagram of the labels for the floating debris dataset;

[0056] Figure 4 These are typical sampling images from the model training process;

[0057] Figure 5 The graph shows a comparison of the accuracy of the invented method (MFL-YOLO) and the baseline model (YOLOv11n) across different categories (left) and a comparison of the mAP curves during the training process (right). Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0059] This invention presents a wave-resistant, multi-scale marine floating debris detection method (MFL-YOLO), based on the YOLOv11 architecture. It constructs an efficient, interference-resistant detection model through three key optimization modules and supporting data augmentation strategies. The overall technical approach consists of three stages: dataset construction, model optimization and training, and mobile terminal deployment and real-time inference, as detailed below:

[0060] 1. Construction of Anti-interference Dataset

[0061] Raw sea surface images under various sea conditions (strong light, broken waves, turbid water, etc.) were collected using unmanned surface vessels and drones. CVAT tools were then used for manual cleaning and annotation to construct a dataset containing seven typical types of marine debris, including plastic bottles, plastic bags, discarded fishing gear, and other fragments. To enhance the model's robustness to interference, a triple data augmentation strategy was employed: Mosaic stitching to enrich the background semantics, HSV illumination transformation to simulate extreme lighting environments such as strong reflections and low light, and Mixup enhancement to improve the model's robustness to noise, effectively mitigating the problem of imbalanced samples.

[0062] 2. Model Architecture Optimization

[0063] This invention optimizes the YOLOv11 architecture in three key dimensions, constructing the MFL-YOLO model. The overall network architecture is as follows: Figure 1 As shown, the specific optimization modules are as follows:

[0064] (1) WIoU v3 loss function based on dynamic non-monotonic focusing mechanism

[0065] The CIoU bounding box regression loss in the traditional detection model is replaced with the WIoU v3 loss function. The WIoU v3 loss function is used to optimize the bounding box regression, and it adaptively adjusts the sample gradient gain through a distance attention mechanism and a dynamic non-monotonic focusing mechanism. This implementation process consists of two stages: first, a basic regression based on distance attention, and second, dynamic focusing based on outlier characteristics.

[0066] In the basic regression stage, the unstable aspect ratio penalty term in CIoU is abandoned. A distance attention mechanism is constructed based on the Euclidean distance between the center points of the predicted box and the ground truth box to adapt to the irregular deformation of floating objects on the sea surface caused by wave fluctuations. In the dynamic focusing stage, a non-monotonic gradient gain function is constructed by calculating the sample outlier degree (the ratio of the current sample loss value to the batch average loss value) and dynamically allocating gradient weights. The weights of simple samples and extremely low-quality samples (wave noise, mislabeling) are reduced, while the learning weights of medium-difficulty samples (blurred, partially occluded targets) are strengthened to block the propagation of harmful gradients.

[0067] Mathematical model: Let the predicted bounding box be B, and the true bounding box be B. gt Basic loss Distance attention coefficient With IoU loss The product is constructed using the following formula:

[0068]

[0069] in, =1-IoU, where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; The effect of center point deviation is amplified by an exponential function, as shown in the formula:

[0070]

[0071] In the formula, (x,y), ( ) are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. g H g Let the width and height be the minimum bounding rectangle of the two.

[0072] Define outlier Used to measure sample quality:

[0073]

[0074] in This represents the average IoU loss for the current batch, and it blocks... Gradient backpropagation; based on Construct the non-monotonic gradient gain coefficient r, and the final loss function is:

[0075]

[0076]

[0077] This formula clearly shows that gradient gain It's about outlier The function is subject to hyperparameters And δ control. The function exhibits a bell-shaped curve distribution: when Smaller or At its maximum, The value is relatively small; while when When in the middle range, The value increases significantly. This non-monotonic weighting mechanism constitutes the core advantage of this invention, enabling adaptive hierarchical processing of complex sea surface samples.

[0078] (2) ASFF noise reduction and fusion module based on dynamic weight remapping

[0079] An ASFF adaptive feature fusion module is introduced into the neck of the model to replace the traditional feature fusion operation and solve the problem of conflict between high-frequency wave noise and multi-scale features on the sea surface.

[0080] An ASFF module is constructed in the neck network to replace the traditional fusion operation. This module first adjusts the features of the shallow (P3), middle (P4), and deep (P5) layers to the same spatial resolution through upsampling or downsampling operations. Then, the module generates a spatial weight map through network self-learning; that is, for each pixel in the image, it automatically calculates the weight coefficients α, β, and γ from the three layers P3, P4, and P5, ensuring that the sum of the three is 1. Finally, these weights are used to perform a weighted summation of the features from the three layers, outputting the fused features. Through this mechanism, when the network detects high-frequency wave reflection noise in the shallow features, it can automatically reduce the weight of the shallow features in that region and instead utilize more deep semantic features for judgment, thereby effectively suppressing sea surface background noise while preserving details of small targets.

[0081] Mathematical model: Let , , The feature vectors of P3, P4, and P5 are adjusted to the resolution of the l-th layer, and the three spatial weight parameters learned at pixel position (i, j) are respectively... , , To ensure that the sum of the weights is 1 and that the values ​​are stable, the network uses the Softmax function to control the parameters. Normalization is performed. The calculation formula is as follows:

[0082]

[0083] In the formula, , , These are the control parameters learned through backpropagation of the network. Similarly, we can obtain... and The final fusion features The calculation is a weighted sum of features from each layer:

[0084]

[0085] This formula shows that the weights for feature fusion are dynamically generated for each pixel in the image. This pixel-level weighting mechanism allows the model to adaptively suppress background noise caused by reflections from nearby waves while preserving features of small distant targets.

[0086] (3) Task-aware dual-path decoupled detection head

[0087] To address the inherent conflict in feature requirements between the classification task of distinguishing between waves and debris and the regression task of accurately locating boundaries in marine floating object detection, as well as the optimization contradiction between locating small targets on the sea surface and suppressing background texture, this invention designs a task-aware dual-path decoupled detection head. Unlike the coupled architecture of existing technologies where classification and regression share feature extraction, this mechanism adopts a task decoupling strategy, completely blocking gradient interference between the two tasks. Furthermore, considering the unique characteristics of the marine environment, a dual-stream context enhancement unit is innovatively constructed within the decoupling branch. This unit, through a parallel vision-aware mechanism, can adaptively balance the information weights of local details and global context according to the current task requirements, thereby significantly reducing the false alarm rate of waves while ensuring the accuracy of irregular floating object detection boxes.

[0088] Specifically, the traditional coupled detection head is replaced with a task decoupling architecture, splitting the input feature stream into independent classification and regression branches. To enhance the expressive power of the features, this invention constructs parallel local and global context-aware units within each decoupling branch. The local feature extraction branch uses standard convolution to extract local geometric features, focusing on capturing the edge and contour details of floating objects. The global context-aware branch uses dilated convolutions with a dilation rate of 2 to expand the receptive field, focusing on perceiving the texture and environmental context of the surrounding sea surface. These two feature streams are fused after extraction and then passed through a task-aware attention module based on sigmoid activation. This module acts as an intelligent gating system, automatically identifying and filtering irrelevant features according to the specific task requirements of the current branch's classification or regression. For example, in the classification branch, the model automatically enhances global texture features that help distinguish waves; while in the regression branch, it focuses more on local edge features that help correct boundaries.

[0089] The input feature stream is divided into independent classification and regression branches to block gradient interference between tasks. Each branch has parallel local and global vision perception units. The local branch extracts target edges and contour details through 3×3 standard convolutions, while the global branch expands the receptive field through dilated convolutions with a dilation rate of d=2 to perceive the texture context of the sea surface. The two feature streams are fused and then processed by a task-aware attention module activated by Sigmoid. This module adaptively filters irrelevant features. The classification branch strengthens global texture features to distinguish between waves and debris, while the regression branch strengthens local edge features for accurate localization.

[0090] Mathematical model: Let the input features be... The output features are The formulas for feature extraction from local and global branches are as follows:

[0091]

[0092]

[0093] In the formula, It is a 3×3 standard convolution. For dilated convolutions with a dilation rate of d=2, the two feature channels are concatenated and then fused using a 1×1 convolutional layer for dimensionality reduction. :

[0094]

[0095] In the formula, This represents a channel splicing operation. The weights represent the weights of the 1×1 convolution kernel. Finally, the sigmoid activation function is used. Construct a self-attention gating mechanism. Fuse features. After being weighted by its own Sigmoid activation value, it is compared with the original input. By performing residual superposition, the final output is obtained:

[0096]

[0097] This formula clearly shows that the output feature consists of two parts: feature extraction and fusion, and the input residual. Among them, the term... This constitutes a task-aware mechanism, which automatically suppresses background noise responses in the fused features that are irrelevant to the current task by using nonlinear activation.

[0098] 3. Model Training and Deployment Inference

[0099] The MFL-YOLO model was trained using a constructed dataset, with WIoU v3 as the loss function during training to iteratively optimize the model parameters. After training, the model format was converted and deployed to mobile terminals such as unmanned surface vessels and drones. In the real-time inference stage, the input sea surface image was processed by the model, and redundant detection boxes were removed using a non-maximum suppression algorithm to output the category and location coordinates of floating debris, achieving real-time monitoring with wave interference resistance.

[0100] Experimental Validation - Dataset Construction and Scene Characteristic Analysis

[0101] To construct a detection model with high generalization ability, this invention collected and created a dataset of marine floating debris covering various sea conditions. (Appendix) Figure 3 The statistical distribution characteristics of the labels in this dataset are shown. The category histogram in the upper left corner reveals a significant sample imbalance, with "Plastic Bottle" and "Misc Debris" being more numerous than categories like "Plastic Bag." This long-tailed distribution validates the necessity of introducing the WIoU loss function to dynamically balance the weights of easy and difficult samples in this invention. Furthermore, the scatter plot in the lower right corner and the bounding box size distribution in the upper right corner show that the vast majority of targets are clustered near the origin in the lower left corner, indicating the presence of a large number of extremely small floating targets in the dataset. This places extremely high demands on the detector's multi-scale feature fusion capabilities.

[0102] Appendix Figure 4 The image showcases typical sampling batches during model training, intuitively reflecting the real-world marine operating environment addressed by this invention. As can be seen from the image, the sampling scenarios encompass strong light reflection, white foam interference from wave breakers, turbid water, and water surface textures under varying lighting conditions. Some targets are barely visible due to wave obscuring, and the background contains a large amount of wave noise resembling white debris. This highly complex environmental characteristic fully demonstrates the practical application value of the ASFF-based noise-resistant fusion mechanism and task-aware decoupling detection head designed in this invention, aiming to accurately extract effective target features from strong background noise.

[0103] To verify the effectiveness of the proposed wave-resistant multi-scale marine debris detection method (MFL-YOLO) under real and complex sea conditions, we constructed an experimental dataset containing seven typical types of marine debris and rigorously compared the proposed method (MFL-WIoU) with the baseline model YOLOv11n. (See attached...) Figure 5The comparison results of detection accuracy across categories shown on the left demonstrate that the method of this invention outperforms the baseline model in terms of mAP50 across all seven test categories. Particularly noteworthy are the significant improvements in detection accuracy for the highly challenging categories of "Fishing Gear" and "Plastic Bag." This result strongly proves the crucial role played by the WIoU loss function introduced in this invention: Faced with low-quality samples such as fishing gear with fine lines or plastic bags susceptible to wave deformation, the dynamic non-monotonic focusing mechanism of WIoU successfully guides the model to give higher attention to these difficult samples during training, thus effectively overcoming the missed detection problem that is prone to occur with traditional loss functions.

[0104] In addition, Figure 5 (Right) This image shows the mAP50 curve of the model during 150 epochs of training, further revealing the advantages of this invention in terms of convergence speed and stability. The curve trend in the early training phase (first 20-40 epochs) shows that the accuracy of the method in this invention increases significantly faster than the baseline model. This indicates that the ASFF adaptive feature fusion mechanism in the neck network accelerates the network's extraction and integration of effective features. As training progresses, the accuracy curve of the method in this invention remains stable above the baseline and eventually converges to an mAP50 level of approximately 0.68, a significant improvement compared to the baseline's 0.61. This phenomenon of maintaining excellent generalization ability without overfitting even after long-term training fully verifies the effectiveness of the task-aware dual-path decoupled detection head in resolving conflicts between classification and regression features, ensuring that the model possesses both high recall and high positioning accuracy in complex sea conditions.

[0105] In summary, this invention can accurately identify marine debris under complex sea conditions such as strong light reflection and wave breaking. Its detection accuracy and anti-interference ability are significantly better than the baseline model. It can be deployed on mobile terminals such as unmanned ships and drones to achieve real-time and efficient monitoring.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scale marine floating debris detection method resistant to wave interference, characterized in that: Includes the following steps: S1. Constructing an interference-resistant marine floating debris dataset: Raw sea surface images under different sea conditions were collected by unmanned vessels and drones. After manual cleaning and CVAT annotation, a dataset containing seven types of floating debris was obtained. A triple strategy of Mosaic stitching, HSV illumination transformation, and Mixup hybrid enhancement was adopted to expand the dataset and improve the noise resistance and robustness of the model. S2. Construct the MFL-YOLO optimization model: Based on YOLOv11, integrate three major modules: (1) The WIoU v3 loss function is adopted, and the sample gradient gain is adaptively adjusted based on distance attention and dynamic non-monotonic focusing mechanism; (2) An ASFF adaptive feature fusion module is embedded in the neck to achieve pixel-level dynamic weight remapping, suppress wave noise and eliminate multi-scale feature conflicts; (3) Design a task-aware dual-path decoupled detection head, combined with a dual-stream context enhancement unit, to block gradient interference in classification and regression tasks; S3. Model Training: Train the MFL-YOLO model based on the constructed dataset, iteratively optimize the parameters using WIoU v3 as the loss function, and save the optimal model weights; S4. Mobile Terminal Deployment and Real-time Inference: After converting the trained model format, deploy it to the mobile terminal of the unmanned vessel / drone to collect sea surface images in real time and input them into the model. Redundant detection boxes are removed through non-maximum suppression algorithm, and the category and location coordinates of floating debris are output to achieve real-time monitoring against wave interference.

2. The method for detecting multi-scale marine floating debris with wave interference according to claim 1, characterized in that: The calculation process of the WIoU v3 loss function includes: a. Constructing a base loss function based on distance attention : In the formula, =1-IoU,(x,y),( ) represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively, W g H g Let the width and height be the minimum bounding rectangle; b. Define outlier β to measure sample quality: In the formula, The regression loss for the current sample; The average IoU loss for the current batch, and The calculation process was interrupted. The gradient is backpropagated so that it is used only as a reference coefficient for the dynamic weights; a non-monotonic gradient gain coefficient r is constructed to obtain the final loss function: In the formula, α and δ are hyperparameters. This is the final loss function.

3. The method for detecting multi-scale marine floating debris with wave interference according to claim 1, characterized in that: The implementation process of the ASFF adaptive feature fusion module includes: a. Unify the features of the three layers of shallow layer P3, middle layer P4 and deep layer P5 to the same spatial resolution by upsampling or downsampling, and calculate the weight coefficients α, β and γ of the three layers P3, P4 and P5, and α+β+γ=1. , , These represent feature vectors from the three levels P3, P4, and P5, respectively, adjusted to the resolution of the current level l. b. Learn pixel-level weight coefficients by normalizing using the Softmax function. , , : In the formula, , , These are the control parameters learned through backpropagation of the network; similarly, we can obtain... and c. Fusion features The calculation is a weighted sum of features from each layer: The weights for feature fusion are dynamically generated for each pixel in the image, adaptively suppressing background noise caused by near-field wave reflections.

4. The method for detecting multi-scale marine floating debris with wave interference as described in claim 1, characterized in that: The implementation process of the task-aware dual-path decoupled detection head includes: a. dividing the input feature stream into independent classification and regression branches, extracting local and global features in parallel within each branch, and using a 3×3 standard convolution function to extract edge detail features in the local branch. The global branch uses a dilated convolution function with a dilation rate of d=2 to extract environmental context features. The mathematical expression is as follows: b. Concatenate the two feature streams along the channel dimension and apply a weight matrix. Convolutional layers are used for dimensionality reduction and feature interaction to generate fused features. : In the formula, This represents a channel splicing operation. c. Weight parameters representing a 1×1 convolution kernel; d. Activated by the Sigmoid function. Construct a self-attention gating mechanism and fuse features After being weighted by its own Sigmoid activation value, it is compared with the original input. By performing residual superposition, the final output is obtained: In the formula, A task-aware mechanism is constructed, which automatically suppresses background noise responses in the fused features that are irrelevant to the current task by using nonlinear activation.

5. The method for detecting multi-scale marine floating debris with wave interference according to claim 1, characterized in that: The seven types of floating waste include plastic bottles, plastic bags, foam, fishing nets, glass bottles, aluminum cans, and other fragments. The dataset is divided into training set:validation set:test set = 7:2:1, and the sample size of the training set is increased to 4 times the original size after augmentation.

6. The method for detecting multi-scale marine floating debris with wave interference according to claim 1, characterized in that: The model training parameters were set as follows: batch size 16, initial learning rate 0.001, cosine annealing learning rate scheduling, weight decay coefficient 0.0005, training epochs 150, confidence threshold for non-maximum suppression algorithm 0.5, and IOU threshold 0.

45.

7. The method according to any one of claims 1-6, characterized in that, The method is deployed on unmanned ships and drone mobile terminals.

Images