Underwater target real-time detection method based on YOLO11s-UW model

Abstract

This invention relates to a real-time underwater target detection method based on the YOLO11s-UW model, belonging to the field of computer vision technology. The method constructs a YOLO11s-UW model to identify underwater targets. This model uses YOLO11s as a baseline model, replacing all but the first Conv module in its backbone and neck networks with the MADConv module. In the backbone network, the GSSDFP module replaces the original SPPF module, while the detection head retains its original structure. The YOLO11s-UW model of this invention performs excellently in underwater biological and plastic waste detection tasks and can be implemented in real-time on edge devices.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and relates to a real-time underwater target detection method based on the YOLO11s-UW model, which is suitable for detecting underwater organisms and plastic waste. Background Technology

[0002] The ocean covers most of the Earth's surface and is a vital component of the Earth's ecosystem. Furthermore, the ocean contains abundant aquatic resources such as oil, natural gas, minerals, and chemicals, thus driving the development of marine exploration activities. Supporting marine exploration requires various underwater operations, including target location, biometrics, underwater archaeology, object recovery, environmental monitoring, and equipment maintenance. Due to the unique and complex characteristics of the underwater environment, conducting these operations solely through human intervention faces numerous challenges.

[0003] Related research data shows that massive amounts of plastic waste enter the marine ecosystem every year, and this figure is increasing annually. Therefore, there is an urgent need to intensify efforts to clean up marine plastic waste. Most plastic waste in oceans and lakes is underwater, making manual cleanup not only difficult but also posing safety hazards. A more efficient and reliable solution is to use intelligent grasping robots. This method is more scientifically sound and is gradually becoming the preferred solution for waste cleanup. For underwater robots to grasp plastic waste, the primary prerequisite is to achieve accurate target identification and positioning. This requires the deployment of high-performance target detection algorithms to ensure the accurate execution of waste collection tasks. Therefore, research on underwater target detection for marine plastic waste is particularly important.

[0004] The rapid development of computer vision technology has broken through the limitations of land applications and begun to address the challenges of underwater scenarios. As a core technology, underwater target detection encompasses marine geodesy, underwater archaeology, explosive ordnance disposal, and fisheries resource management. Underwater target detection is mainly divided into two categories: acoustic-based detection systems and optical-based detection systems. Compared to acoustic images, optical images have higher resolution and richer information content, and their acquisition methods are also more economical. Therefore, research on optical-based underwater target detection has significant practical value and scientific significance.

[0005] Since the advent of object detection algorithms in 1998, traditional methods have undergone significant evolution, giving rise to numerous research schemes centered on hand-designed features. Underwater object detection and recognition systems based on classical features rely on manually designed target features, typically employing low-level image descriptors for detection. These methods exhibit robust performance when detecting targets with distinct visual features, but their performance deteriorates in complex scenes and they are highly susceptible to environmental interference. Furthermore, traditional methods require manual feature selection and extraction, while deep learning methods effectively address these challenges. Deep learning-based object detection algorithms are mainly divided into two-stage and single-stage algorithms. Currently, in practical applications of real-time object detection, single-stage detection algorithms are commonly used, primarily represented by the YOLO series algorithms, including YOLOv5 and YOLOv8. Single-stage object detection frameworks skip the region proposal generation step, treating the detection task as an end-to-end regression problem. Compared to two-stage methods, these methods offer superior computational efficiency, making them particularly suitable for real-time underwater biological target detection scenarios on intelligent platforms such as autonomous underwater vehicles (AUVs).

[0006] In recent years, many researchers have conducted research on underwater real-time target detection algorithms and made some progress. Zhai et al. improved YOLOv5s for underwater sea cucumber detection tasks by introducing a multi-scale Retinex enhancement algorithm combined with color restoration in the image preprocessing stage, integrating a convolutional attention module, and adding a detection layer to the head network. Validation results on a custom dataset show that the improved model achieves a 9% improvement in accuracy and an 11.5% improvement in recall compared to YOLOv5s, fully demonstrating its effectiveness in aquatic organism identification tasks.

[0007] Lei et al. optimized YOLOv5s by replacing the original backbone network with a Swing Transformer to improve feature representation capabilities, optimizing the multi-resolution feature fusion strategy to enhance multi-scale detection performance, and improving the confidence loss function to increase localization accuracy. Test results on a custom dataset built on URPC (containing 6034 images of four aquatic organisms: sea cucumbers, sea urchins, scallops, and starfish) show that the improved model achieves a mean average precision (mAP) of 87.2%, outperforming benchmark algorithms such as SSD, YOLOv4, YOLOv5s, and YOLOv5x.

[0008] Sun et al. proposed a lightweight YOLOv8s variant. They replaced the original backbone network with a FasterNet module and optimized the feature pyramid. They also replaced the final standard convolutions and the bottleneck C2f structure with GSConv (a variant of Ghost Shuffle Convolution or Generic Shuffle Convolution) and partial convolutions, constructing a lightweight C2f architecture. This model achieved mAP50 of 86.8%, 84.3%, and 84.7% on the RUOD, UTDAC2020, and URPC2022 datasets, respectively, surpassing YOLOv8s in performance while reducing the model size by 24%.

[0009] Zhou et al. proposed the UODN algorithm based on YOLOv8. It addresses the challenges of poor underwater image quality and multi-scale target detection through a Cross-Stage Multi-Branch (CSMB) module and a Large Kernel Spatial Pyramid (LKSP) module: the CSMB module enhances feature extraction capabilities for low-quality images, while the LKSP module improves multi-scale detection performance. Experiments on the URPC2020 dataset show that, compared to YOLOv8s, UODN improves AP by 1.0%, AP50 by 1.1%, and AP75 by 2.1%.

[0010] Lu et al. proposed the SCoralDet algorithm based on YOLO for underwater soft coral detection. This algorithm addresses the challenges of low image contrast and complex coral structures through Multi-Path Fusion Block (MPFB) and lightweight components: the MPFB module enhances the robustness of multi-scale features, while the lightweight components achieve a balance between performance and efficiency. On the Soft-Coral dataset, the model achieves an mAP50 of 81.9 and an inference latency of 9.52 milliseconds, outperforming mainstream YOLO variants.

[0011] While researchers have proposed several excellent detection models for real-time underwater target detection, their detection metrics are not particularly high, leaving room for improvement. Furthermore, many past models were improvements upon existing frameworks of YOLOv5 and YOLOv8. On September 30, 2024, YOLOv11 was released online, featuring significant improvements in architecture and training methods. It boasts enhanced feature extraction, optimized efficiency and speed, higher accuracy with fewer parameters, and can be deployed across various environments, supporting a wide range of computer vision tasks. Using the publicly available underwater target detection dataset RUOD for training, preliminary experiments showed that the YOLOv11 model still has room for improvement in its detection results on underwater target detection datasets. For some targets clearly visible to the human eye in certain images, the model failed to identify them correctly. Therefore, improvements based on the YOLOv11 framework are considered to maximize its detection accuracy on underwater target detection data. Summary of the Invention

[0012] In view of this, the purpose of this invention is to provide a real-time underwater target detection method based on the YOLO11s-UW model, which solves the problem that existing underwater target detection models are difficult to balance between accuracy, computational efficiency and real-time performance, and have poor adaptability to complex underwater scenarios. This method is suitable for detecting underwater organisms and plastic waste.

[0013] To achieve the above objectives, the present invention provides the following technical solution: A real-time underwater target detection method based on the YOLO11s-UW model includes the following steps: S1: Construct a high-quality underwater target detection image dataset and divide it into training set, validation set and test set; S2: Constructing the YOLO11s-UW model: Using YOLO11s as the baseline model, in the backbone network and neck region of the YOLO11s baseline model, the first Conv (Conv2d+BN+SiLU) module is retained, and all other Conv modules are replaced with MADConv (Multi-branch Attention-Driven Conv) modules. In the backbone network, the original SPPF module is replaced with GSSDFP (GSConv Shared Dilated Feature Pyramid combined with GSConv). The detection head maintains the original structure, forming a detection model that balances feature extraction capability and computational efficiency. S3: Train the YOLO11s-UW model using the training set to obtain the optimal parameters of the model; S4: Input the image of the target to be detected into the trained YOLO11s-UW model to identify the target.

[0014] Furthermore, in step S1, a high-quality underwater target detection image dataset is constructed, which specifically includes: firstly, obtaining the open-source underwater biological detection dataset RUOD and the underwater plastic waste detection dataset DeepTrash; then, combining the two datasets and checking and correcting errors to obtain the RUODPLUS dataset, which is then divided into training set, validation set and test set.

[0015] Furthermore, in step S2, the MADConv module includes a feature extraction component, a channel attention module, and a weighted hybrid dynamic spatial attention (WHDSA) module.

[0016] The workflow of the MADConv module is as follows: 1) Feature extraction component: (1) The input feature map is processed by a 1×1 convolutional block (containing Conv2d, batch normalization BN, and SiLU activation function) to achieve downsampling and feature enhancement by doubling the number of channels and halving the height and width of the feature map; (2) Multi-scale features are extracted by two parallel deep dilated convolution (DWConv) branches (dilation rates of 1 and 2, respectively). The branch outputs are spliced ​​along the channel dimension and then nonlinear transformation is introduced by batch normalized BN layer and SiLU activation function. (3) The number of channels is adjusted to twice the number of input channels by using 1×1 convolutional blocks to complete the channel dimension feature fusion; 2) The features are sequentially weighted and optimized by the channel attention module and the WHDSA module to output the final feature map.

[0017] Furthermore, in step S2, the channel attention module is an improvement based on CBAM (Convolutional Block Attention Module), which extracts and fuses channel features through average pooling and max pooling to enhance the weights of key channels.

[0018] Furthermore, in step S2, the WHDSA module includes two branches: a traditional spatial attention branch and a dynamic spatial attention branch based on the former and incorporating DSA (dynamic spatial attention mechanism); the attention maps generated by the two branches are combined through dynamic weighted fusion.

[0019] The WHDSA module cleverly integrates dynamic spatial attention with traditional spatial attention, precisely focusing on the target spatial region by dynamically generating convolution kernels and adaptive weighting.

[0020] Furthermore, in step S2, the GSSDFP module includes a GSConv module, a multi-branch shared weight dilated convolutional structure, and a feature fusion component; The workflow of the GSSDFP module is as follows: (1) The input feature map is optimized by the GSConv module to achieve lightweight feature extraction; (2) A multi-branch shared weight dilated convolution structure is adopted, in which each branch reuses the same set of 3×3 convolution kernels with dilation rates of 1, 3 and 5, respectively, covering receptive fields of different scales, taking into account both local details and global context; (3) Feature fusion component: After the outputs of each branch are spliced ​​along the channel dimension, feature fusion is completed by 1×1 convolution, and multi-scale fused features are output.

[0021] Among them, by reducing parameter redundancy through weight sharing and combining the lightweight characteristics of GSConv, multi-scale detection capabilities are improved without increasing computational costs.

[0022] Furthermore, in step S2, the GSConv module includes standard convolutional blocks, depthwise convolution, and channel rearrangement operations; wherein, the first Conv block includes 1×1 convolution operations, batch normalization, and operations to process the feature map using the SiLU activation function; the depthwise convolution part includes 5×5 convolution operations and batch normalization operations.

[0023] The beneficial effects of this invention are as follows: 1) Significantly Improved Detection Accuracy: This solution constructs a YOLO11s-UW model, including the MADConv and GSSDFP modules. The MADConv network module enhances the underwater target feature extraction capability, while the GSSDFP network module optimizes multi-scale feature fusion. On the RUODPLUS underwater target detection dataset, which includes various underwater environments, the model achieves an mAP50 of 91.5% and an mAP50-95 of 67.2%, respectively, which are improvements of 1.5% and 1.0% compared to YOLO11s, and outperform existing mainstream underwater detection models.

[0024] 2) Excellent computational efficiency and real-time performance: The lightweight design of this solution results in a model parameter count of only 13.8M and a computational load of 32.0GFLOPs, which is far lower than similar high-performance models. The detection speed is approximately 120FPS, meeting the real-time detection requirements of edge devices.

[0025] 3) Strong environmental adaptability: The module is optimized to address issues such as underwater lighting changes, low contrast, and multi-scale targets, making it adaptable to complex underwater scenarios.

[0026] 4) High practicality: It has low computational load, high accuracy, and is suitable for deployment on existing AUV platforms. It does not require a large amount of hardware costs and provides technical support for scenarios such as marine debris cleanup and biological monitoring. Its practical value and promotion prospects are outstanding.

[0027] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the real-time underwater target detection method based on the YOLO11s-UW model of the present invention; Figure 2 This is a structural diagram of the YOLO11s-UW model; Figure 3 Here is a structural diagram of the MADConv module; Figure 4 Here is a structural diagram of the channel attention module; Figure 5 Here is a structural diagram of the WHDSA module; Figure 6 Here is a structural diagram of the DSSDFP module; Figure 7 Here is a structural diagram of the GSConv module; Figure 8 A comparison chart of the mAP50 index of different detection models on the RUOD dataset. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] Please see Figure 1 This invention provides a real-time underwater target detection method based on the YOLO11s-UW model, the overall process of which mainly includes the following key steps: Step 1: Construct a high-quality underwater target detection dataset.

[0031] First, we acquired the open-source underwater biological detection dataset RUOD and the underwater plastic waste detection dataset DeepTrash. Then, we combined the two datasets, checked and corrected errors, to obtain the RUODPLUS dataset, which was then divided into training, validation, and test sets. RUODPLUS contains nearly 20,000 underwater images from various environments, covering 13 categories: 11 biological categories and 2 plastic waste categories. After the dataset was divided, the training set contains 13,327 images, the validation set contains 3,228 images, and the test set contains 3,249 images.

[0032] Step 2: Use the dataset to train the YOLO11s model.

[0033] Training Environment: This example uses an AutoDL cloud server for the experiment. The server configuration is as follows: NVIDIA RTX4090D (24GB) graphics card, 16-core Intel(R) Xeon(R) Platinum 8481C processor, 80GB of memory, 80GB of storage space, and the operating system is Ubuntu 22.04. The experimental environment is built based on PyTorch 2.3.0, Python 3.12, and CUDA 12.1.

[0034] Training Strategy: Caching was not enabled during model training; the input image size was set to 640×640 pixels, and the batch size was set to 48 to ensure the model could learn fully from large-scale data. Regarding the training strategy, 300 training epochs were used. Mosaic data augmentation was disabled in the 10th epoch to prevent the model from over-adapting to augmentation effects and ignoring original data features in the later stages of training. Stochastic Gradient Descent (SGD) was used as the model optimizer, and Automatic Mixed Precision (AMP) training was enabled to effectively reduce GPU memory usage. All other hyperparameters used the default configurations from the official YOLO11 project. Subsequent YOLO11s-UW model training also adopted the same training configuration.

[0035] Deployment scenarios: Suitable for edge devices such as autonomous underwater vehicles (AUVs), supporting real-time tasks such as underwater target localization, biometric identification, and plastic waste cleanup.

[0036] Step 3: Construct the YOLO11s-UW model, its network structure is as follows: Figure 2 As shown.

[0037] The YOLO11s-UW utilizes the MADConv module and GSSDFP module designed in this invention. The structure and workflow of these two modules are described below.

[0038] The structure diagram and workflow of the MADConv module are as follows: Figure 3 As shown. The MADConv module mainly consists of three components: a dual-branch feature extraction component, a channel attention module, and a weighted hybrid dynamic spatial attention (WHDSA) module. The input feature map is denoted as... Where B represents the batch size, C is the number of channels, and H and W represent the height and width of the feature map, respectively.

[0039] First, the input feature tensor is fed into a 1×1 Conv module (Conv2d+BN+SiLU). This module doubles the number of channels while halving both the height and width of the feature tensor: halving the height and width simulates the downsampling effect applied to the input feature map by the Conv module in the YOLO11 backbone network; while doubling the number of channels allows the model to capture richer feature information, including some spatial features that might be lost when the height and width are reduced.

[0040] Next, the feature map is fed into two parallel deep convolutional branches with dilatations of 1 and 2, respectively, to extract features at different scales. Then, the outputs of the two parallel deep branches are concatenated along the channel dimension, at which point the number of channels in the feature map becomes 4C, further enriching the feature information. Next, the concatenated feature tensor is processed by a batch normalization (BN) layer, and after the BN step, it is fed into a SiLU activation function to introduce non-linearity.

[0041] Subsequently, a Conv module with a 1×1 kernel is used to adjust the number of channels to 2C. This convolutional block performs pointwise convolution: fusing feature tensors along the channel dimension, thereby enhancing the capture of channel-dimensional feature representations. Finally, the feature map is processed sequentially by the channel attention module and the WHDSA module to obtain the final output features of the MADConv module. Its calculation process is shown in Equation 1-3.

[0042] (1) (2) (3) Where F represents the input feature tensor; The symbol represents a two-dimensional convolution operation; c1 and c2 represent the width and height of the convolution kernel, respectively; Concat represents the concatenation function; BN represents batch normalization; and SiLU represents the activation function.

[0043] The network structure of the channel attention module is as follows: Figure 4As shown in the figure. The channel attention module used in MADConv is derived from the channel attention module of CBAM, with slight modifications, including replacing the original fully connected layer with a convolution operation and using SiLU activation function. Its calculation process is shown in Equation 4-7.

[0044] (4) (5) (6) (7) The WHDSA module combines the advantages of DSA attention mechanisms with traditional spatial attention mechanisms, enabling it to more accurately focus on the spatial importance of features, such as... Figure 5 As shown, this module comprises two branches: a traditional spatial attention branch and a dynamic spatial attention branch based on the former and incorporating DSA. The attention maps generated by the two branches are combined through dynamic weighted fusion.

[0045] Assume the tensor input to the WHDSA module is This module first applies global average pooling (GAP) to F, compressing it into a global feature tensor; then, this feature tensor is fed into a dynamic kernel generator, first passing through a Conv2d layer to adjust the number of channels. (Followed by the SiLU activation function), then another Conv2d layer is used to prune the number of channels to 9 (followed by the Sigmoid activation function); the output tensor is reshaped into B×1×3×3, where the 3×3 matrix in each sample of this tensor is later used as the convolution kernel parameters. The calculation process of the convolution kernel weight parameters is shown in Equation 8.

[0046] (8) in, This represents a batch of 3×3 convolutional kernels (one for each sample in the batch), which are dynamically generated based on the input F.

[0047] The input feature tensor F is subjected to average pooling and max pooling along the channel dimension to obtain the tensor. and .Will The sample is reshaped into a 1×B×H×W shape, and then a grouped convolution operation with B groups is performed with the previously generated dedicated convolution kernel to ensure that each sample is processed by its own dedicated convolution kernel; the result is reshaped back into a B×1×H×W shape, denoted as […]. .then, and The channels are concatenated along the channel axis, and then a 7×7 Conv2d layer is applied to reduce the number of channels to 1. Finally, the dynamic spatial attention weights are obtained through a Sigmoid activation function. The calculation process is shown in Formula 9-10.

[0048] (9) (10) also, and The tensor is concatenated along the channel dimension, then a 7×7 Conv2d layer is applied to reduce the number of tensor channels to 1. Finally, the traditional spatial attention weights are obtained by applying the Sigmoid activation function. Here, dynamic coefficients are introduced. As learnable parameters. Specifically, we initialize a baseline term. For `nn.Parameter` (a component in PyTorch that can be updated by the optimizer), set it to a small negative value, such as -5 or -10. Then, in the forward propagation method, activate it using the Sigmoid activation function. Map to the interval [0,1].

[0049] (11) Will Multiply by dynamic coefficient Weighted, then combined with Add them together to get the final spatial attention weights. Then input features and Element-wise multiplication is performed to complete the attention weighting process. The calculation process is shown in Formula 12-14.

[0050] (12) (13) (14) This invention proposes the GSSDFP (GSConv SharedDilated Conv Feature Pyramid) module. This module improves multi-scale feature extraction capabilities while maintaining computational efficiency by incorporating lightweight convolutions, parameter-sharing multi-branch dilated convolutions, and multi-scale feature fusion strategies. The specific network structure of the GSSDFP module is as follows: Figure 6 As shown.

[0051] The GSSDFP module embeds the GSConv module (e.g., ...) at the front end of its parallel branch. Figure 7As shown in the diagram, this optimizes the representation of basic features. The GSConv module achieves a balance between accuracy and speed by employing standard convolutional blocks, depthwise convolution, and channel rearrangement operations. The first Conv block includes 1×1 convolution operations, batch normalization, and processing of the feature map using the SiLU activation function. The depthwise convolution part includes 5×5 convolution operations and batch normalization.

[0052] In this parallel multi-convolution operation branch structure, this invention employs parameter-shared dilated convolution. Multiple branches with different dilation rates (1, 3, and 5, respectively) reuse the same set of 3×3 convolution kernels, leveraging dilated convolution to achieve coverage of multi-scale receptive fields. Specifically, low-dilation-rate branches focus on local detail features, while high-dilation-rate branches capture global contextual information. Simultaneously, the parameter-sharing mechanism significantly reduces parameter redundancy, effectively balancing the model's feature representation capability and computational efficiency. Finally, the output of the GSConv block and the outputs of the three branches are concatenated along the channel dimension, and then a 1×1 Conv block (containing Conv2d+BN+SiLU operations) is used to further extract channel-dimensional features and adjust the number of channels. This module, by integrating lightweight convolution, multi-branch parameter-shared dilated convolution, and a multi-scale feature fusion mechanism, effectively improves multi-scale feature extraction performance while maintaining the original computational efficiency.

[0053] Step 4: Evaluate the performance of the proposed underwater target detection model and compare it with other detection models.

[0054] Evaluating the performance of an object detection model requires a comprehensive consideration of detection accuracy, localization precision, and computational efficiency. This section describes the core evaluation metrics used in this experiment. Precision quantifies the proportion of true positive samples among the results identified as positive by the model; this metric reflects the accuracy of the model in predicting positive samples. Recall characterizes the proportion of correctly detected positive samples among all true positive samples, reflecting the model's coverage of positive samples.

[0055]

[0056]

[0057] Average precision (AP) is a metric that measures the accuracy of an object detection model in detecting a single class. It is calculated by calculating the precision at multiple recall thresholds and then averaging the results.

[0058]

[0059] Mean Average Precision (mAP) is the arithmetic mean of the average precision (AP) across all target categories and is a core evaluation metric for multi-category target detection tasks. The calculation of mAP50 requires setting the Intersection over Union (IoU) threshold to 0.5, while the calculation of mAP50-95 uses an IoU threshold that increases in increments of 0.05 from 0.5 to 0.95.

[0060]

[0061] Frames per second (FPS) is used to evaluate the real-time processing capability of a model. It is calculated by counting the number of images that the model can process per second.

[0062]

[0063] Params (short for model parameters) refer to the total number of parameters contained in a model, used to measure the complexity level of the model. GFLOPS (short for Giga Floating Point Operations) is used to evaluate the computational load of a network, characterizing the number of floating-point operations performed by the model.

[0064] To verify the superiority of the proposed YOLO11s-UW model in real-time underwater target detection tasks, its detection performance was compared with that of several advanced and representative target detection models and a few high-performing underwater target detection models. For YOLO series models, the Ultralytics framework was used for training; while the Faster-RCNN model was trained on the MMDetection platform. Given the structural differences among various detection models, using a uniform hyperparameter configuration across all models would prevent many from achieving optimal performance. Therefore, this invention uses default hyperparameter configurations for training different models, and the training process continues until the algorithm fully converges.

[0065] Comparative experimental results show that the model proposed in this invention achieves optimal performance on the RUODPLUS dataset, with mAP50 and mAP50-95 scores reaching 91.5 and 67.2 respectively, thus outperforming almost all other comparative models, as shown in Table 1. This model is only slightly inferior to the YOLO11m model in the mAP50-95 score, but surpasses the YOLO11m model in mAP50, floating-point operations (GFLOPs), and total number of parameters. Furthermore, the detection speed of this model also meets the requirements of real-time detection. These results fully demonstrate that the model developed in this invention achieves a good balance between detection efficiency and detection accuracy, and is an efficient technical solution suitable for real-time underwater target detection tasks.

[0066] Table 1 Model Performance Comparison

[0067] Furthermore, we trained the model on the open-source underwater target detection dataset RUOD and validated its metrics, comparing it with other detection models. Figure 8 As shown, the YOLO11s-UW model also performs excellently, comparable to the YOLO11m model, but its computational cost is less than half that of the YOLO11m model, demonstrating the significant application potential of the YOLO11s-UW model in resource-constrained environments. The experimental results for the YOLO11n, YOLO11s, YOLO11m, YOLOv8s, and YOLO11s-UW models are all derived from experiments conducted during the research process of this invention. The mAP50 values ​​of the other models on this dataset are derived from data in relevant papers.

[0068] This invention also systematically evaluated the impact of the MADConv and GSSDFP modules on the performance of the YOLO11s-UW model. Using YOLO11s as the baseline model, this invention designed four sets of control experiments: the first set of experiments verified the basic performance of the baseline model; the second set of experiments replaced the convolutional blocks in YOLO11s with the MADConv module to verify its excellent feature extraction capability; the third set of experiments replaced the original SPPF module with the GSSDFP module to evaluate its improvement in multi-scale feature extraction capability; the last set of experiments integrated the above two modules to verify their synergistic optimization effect. The experimental results are detailed in Table 2.

[0069] Table 2 Ablation Experiment Results

[0070] In summary, the YOLO11s-UW model performs excellently in underwater biological and plastic waste detection tasks, can be run in real time on edge devices, and has broad application prospects.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time underwater target detection method based on the YOLO11s-UW model, characterized in that, The method includes the following steps: S1: Construct a high-quality underwater target detection image dataset and divide it into training set, validation set and test set; S2: Constructing the YOLO11s-UW model: Using YOLO11s as the baseline model, in the backbone network and neck region of the YOLO11s baseline model, the first Conv module is retained, and all other Conv modules are replaced with MADConv modules. In the backbone network, the original SPPF module is replaced with the GSSDFP module, and the detection head maintains the original structure. Here, MADConv represents multi-branch attention-driven convolution, and GSSDFP represents dilated convolutional feature pyramid with shared parameters combined with GSConv. S3: Train the YOLO11s-UW model using the training set to obtain the optimal parameters of the model; S4: Input the image of the target to be detected into the trained YOLO11s-UW model to identify the target.

2. The real-time underwater target detection method according to claim 1, characterized in that, In step S1, a high-quality underwater target detection image dataset is constructed, which specifically includes: first, obtaining the open-source underwater biological detection dataset RUOD and the underwater plastic waste detection dataset DeepTrash; then, combining the two datasets and checking and correcting errors to obtain the RUODPLUS dataset, which is then divided into training set, validation set and test set.

3. The real-time underwater target detection method according to claim 1, characterized in that, In step S2, the MADConv module includes a feature extraction component, a channel attention module, and a WHDSA module, wherein WHDSA represents weighted hybrid dynamic spatial attention; The workflow of the MADConv module is as follows: 1) Feature extraction component: (1) The input feature map is processed by a 1×1 convolution block to achieve downsampling and feature enhancement by doubling the number of channels and halving the height and width of the feature map; (2) Multi-scale features are extracted by two parallel depth-separable dilated convolution branches. The branch outputs are spliced ​​along the channel dimension and then nonlinear transformation is introduced by batch normalized BN layer and SiLU activation function. (3) The number of channels is adjusted to twice the number of input channels by using 1×1 convolutional blocks to complete the channel dimension feature fusion; 2) The features are sequentially weighted and optimized by the channel attention module and the WHDSA module to output the final feature map.

4. The real-time underwater target detection method according to claim 3, characterized in that, In step S2, the channel attention module is an improvement based on CBAM. It extracts and fuses channel features through average pooling and max pooling to enhance the weights of key channels. Here, CBAM stands for Convolutional Block Attention Module.

5. The real-time underwater target detection method according to claim 3, characterized in that, In step S2, the WHDSA module includes two branches: a traditional spatial attention branch and a dynamic spatial attention branch based on the former and incorporating DSA, where DSA represents the dynamic spatial attention mechanism; the attention maps generated by the two branches are combined through dynamic weighted fusion.

6. The real-time underwater target detection method according to claim 1, characterized in that, In step S2, the GSSDFP module includes a GSConv module, a multi-branch shared weight dilated convolutional structure, and a feature fusion component; The workflow of the GSSDFP module is as follows: (1) The input feature map is optimized by the GSConv module to achieve lightweight feature extraction; (2) A multi-branch shared weight dilated convolution structure is adopted, in which each branch reuses the same set of 3×3 convolution kernels with dilation rates of 1, 3 and 5, respectively, covering receptive fields of different scales, taking into account both local details and global context; (3) Feature fusion component: After the outputs of each branch are spliced ​​along the channel dimension, feature fusion is completed by 1×1 convolution, and multi-scale fused features are output.

7. The real-time underwater target detection method according to claim 1, characterized in that, In step S2, the GSConv module includes standard convolutional blocks, depthwise convolution, and channel rearrangement operations; wherein, the first Conv block includes 1×1 convolution operations, batch normalization, and operations to process the feature map using the SiLU activation function; the depthwise convolution part includes 5×5 convolution operations and batch normalization operations.

Images