Lightweighting-based underwater target detection method
By improving the lightweight processing of the YOLOv8 model and combining it with RepGhost, SPD convolution, GFPN and CLLA detection heads, the problems of underwater target detection accuracy and model complexity were solved, achieving efficient and accurate underwater target detection.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- GUANGDONG OCEAN UNIVERSITY
- Filing Date
- 2025-11-05
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional underwater target detection methods lack sufficient accuracy in complex underwater environments and fail to effectively consider model lightweighting, leading to difficulties in resource management and ecological monitoring.
The YOLOv8 model is improved by adopting a lightweight RepGhost reparameter module, SPD convolution module, global feature pyramid GFPN, and CLLA detection head to achieve feature reuse, feature fusion, and target recognition, while reducing computational complexity.
It improves the accuracy and robustness of underwater target detection under limited computing resources, adapts to UAV platforms with limited hardware performance, and achieves accurate detection.
Smart Images
Figure CN2025132791_15052026_PF_FP_ABST
Abstract
Description
A lightweight underwater target detection method Technical Field
[0001] This invention relates to the field of underwater target detection technology, and in particular to a lightweight underwater target detection method. Background Technology
[0002] Underwater target detection plays a crucial role in fisheries resource assessment and ecological environmental protection. With increasing global focus on sustainable development, accurate monitoring of the state of underwater ecosystems and resources is particularly important. In recent years, deep learning-based target detection technology has been widely applied in various fields. Deep learning-based target detection algorithms are mainly divided into two-stage and one-stage algorithms. The former generates candidate regions first, then classifies and locates these regions, achieving high detection accuracy, but suffers from complex structure and poor real-time performance. The latter, on the other hand, completes target detection directly in a single forward propagation, without the need to generate candidate regions.
[0003] However, traditional underwater target detection methods often face challenges in complex underwater environments, such as insufficient lighting, clutter interference, and limited computational resources, leading to insufficient detection accuracy and a lack of consideration for lightweight models to meet practical deployment needs. These problems limit effective resource management and ecological monitoring. Therefore, developing efficient and reliable underwater target detection technologies will not only help improve the accuracy of fisheries resource assessment but also provide a scientific basis for ecological environmental protection, ensuring the sustainable development of marine ecosystems. Summary of the Invention
[0004] To address the aforementioned shortcomings of existing technologies, this invention provides a lightweight underwater target detection method that solves the problems of traditional methods being unable to perform accurate detection in complex underwater environments and with limited computing resources, as well as the failure to consider lightweight models.
[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a lightweight underwater target detection method, comprising the following steps:
[0006] S1: Using YOLOv8 as the baseline model, the lightweight RepGhost reparameter module is used to implement feature reuse on the YOLOv8 network backbone, and the SPD convolution module is used to replace the convolutional layers of the network backbone.
[0007] S2: Replace the network neck of the YOLOv8 with the global feature pyramid GFPN and perform feature fusion;
[0008] S3: Replace the YOLOv8 network header with the CLLA detection head to improve target identification and localization, and obtain a lightweight underwater target detection model;
[0009] S4: Evaluate the performance of the lightweight underwater target detection model and apply it to a real-world scenario to achieve underwater target detection.
[0010] Furthermore, in S1, the RepGhost reparameter module utilizes reparameterization technology to sequentially add a batch normalization layer, an Add connection layer, and an activation layer after the deformable convolutional layer, and adds a batch normalization branch between the deformable convolutional layer and the Add connection layer.
[0011] Furthermore, the SPD convolution module in S1 performs the following operations:
[0012] S11: For dimensions of Original feature map Feature sub-maps are obtained by downsampling using a scaling factor. The spatial dimension of the original feature map. The number of channels in the original feature map;
[0013] S12: Connect the feature sub-maps along the channel dimension to obtain the feature map. , This is the proportionality coefficient;
[0014] S13: Add a character with The convolutional layer of the filter converts the feature map Convert to .
[0015] Furthermore, the formula for obtaining the feature sub-map in S11 is:
[0016] ;
[0017] , ;
[0018] in, For feature sub-maps, The feature map or tensor representing the input. and This is the subscript index of the coordinate.
[0019] Furthermore, in S2, the global feature pyramid GFPN utilizes CSPStage to achieve skip connections, while replacing C2f and combining it with convolutional layers.
[0020] Furthermore, the CLLA detection head in S3 includes a CLLA cross-layer local attention mechanism layer and a DFL loss function layer, which are connected by a convolutional layer.
[0021] Furthermore, the CLLA cross-layer local attention mechanism layer performs the following operations:
[0022] S31: Assuming low-level feature maps and intermediate feature maps Contains shallow information, high-level feature maps Contains deep semantic information;
[0023] S32: Employs average pooling and Convolution reduces the size of low-level feature maps. and intermediate feature maps The spatial size is determined, and the number of channels is standardized to obtain a new low-level feature map. and new intermediate feature map Make the size and high-level feature map same;
[0024] S33: Employ learnable parameters to transform the new low-level feature map Projected to In the space, the new intermediate feature map Projected to and In the space, the formula is:
[0025] ;
[0026] ;
[0027] ;
[0028] in, , and These are query, key, and value vectors, respectively. , and This indicates that three learnable parameters are used to generate the query, key, and value vectors, respectively.
[0029] S34: Calculate using dot product and softmax function and The relevant weights, and then with Perform a dot product to obtain a new feature map;
[0030] S35: Add the new feature map to the high-level feature map Obtain aggregated feature map The formula is:
[0031] ;
[0032] in, For activation functions, superscript This is the transpose of the matrix. This is the vector dimension used to calculate the attention weights.
[0033] Furthermore, the DFL loss function layer utilizes the cross-entropy function to focus the network distribution around the label value, as shown in the formula:
[0034] ;
[0035] in, For the predicted value in and Distributed focus loss between two adjacent locations, For the network Sigmoid at position The output, For the network Sigmoid at position The output, For tags In position The probability, For tags In position The probability of.
[0036] The beneficial effects of this invention are as follows: The underwater target detection model LFN-YOLO proposed in this invention can better meet the balance between accuracy and model complexity of underwater target detection models, can adapt to target detection on underwater UAV platforms with limited hardware performance, has better accuracy and size compared with mainstream vision models, and can achieve accurate detection of underwater targets with limited computing resources. Attached Figure Description
[0037] Figure 1 is a flowchart of a lightweight underwater target detection method.
[0038] Figure 2 shows the network structure of the underwater target detection model LFN-YOLO.
[0039] Figure 3 is a schematic diagram of the RepGhost reparameterization process.
[0040] Figure 4 shows when When =2, Space-to-depth transformed feature map.
[0041] Figure 5 shows the characteristic fusion structure of GFPN.
[0042] Figure 6 is a structural diagram of the CLLA cross-layer local attention mechanism layer. Detailed Implementation
[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0044] As shown in Figure 1, a lightweight underwater target detection method includes the following steps:
[0045] S1: Using YOLOv8 as the baseline model, the lightweight RepGhost reparameter module is used to implement feature reuse on the YOLOv8 network backbone, and the SPD convolution module is used to replace the convolutional layers of the network backbone.
[0046] S2: Replace the network neck of the YOLOv8 with the global feature pyramid GFPN and perform feature fusion;
[0047] S3: Replace the YOLOv8 network header with the CLLA detection head to improve target identification and localization, and obtain a lightweight underwater target detection model;
[0048] S4: Evaluate the performance of the lightweight underwater target detection model and apply it to a real-world scenario to achieve underwater target detection.
[0049] This invention proposes an algorithm for underwater target recognition that improves the detection performance of small targets while maintaining the detection performance of normal-sized targets, and reduces the number of model parameters and computational complexity. As shown in Figure 2, firstly, this invention proposes to improve the YOLOv8 network backbone by using the RepGhost network structure reparameterization module to achieve feature reuse. Secondly, it replaces the progressive convolution with the SPD convolution module in the network backbone and adds a pooling layer to the network. Then, the neck feature fusion method is improved to GFPN, enabling sufficient information exchange between high-level semantic information and low-level spatial information. Finally, a CLLA detection head combining cross-layer attention mechanism and distributed focus loss is proposed to replace the original YOLOv8 detection head, thereby improving the model's target recognition and localization, ultimately forming the LFN-YOLO underwater target detection model.
[0050] The RepGhost reparameter module in S1 uses reparameterization technology to add a batch normalization layer, an Add connection layer, and an activation layer sequentially after the deformable convolutional layer, and adds a batch normalization branch between the deformable convolutional layer and the Add connection layer.
[0051] Feature reuse plays a crucial role in lightweight convolutional neural networks. Existing feature reuse methods typically utilize concatenation operations to reuse feature maps from other layers to maintain a large number of channels. However, this approach incurs significant computational costs on hardware devices, posing challenges for practical industrial applications. Therefore, this solution proposes incorporating the RepGhost module into the YOLOv8 network backbone to achieve feature reuse through structural reparameterization, thus eliminating the computational burden caused by the concatenation operators commonly used in previous feature reuse methods.
[0052] The RepGhost reparameterization module replaces the concatenation operator in Ghost with the Add operator, which is faster and more efficient. Then, the ReLU activation layer is moved after the deep convolutional layer and the addition layer to meet the rules of the reparameterization structure. Finally, a batch normalization (BN) branch is added to the module, and deep convolutional layers are fused during inference to reduce floating-point operations. Figure 3 illustrates the RepGhost reparameterization process.
[0053] Introducing the RepGhost module into the YOLOv8 network backbone allows for more efficient training of object detection models. It effectively improves detection speed during the inference phase while minimizing accuracy loss, balancing simplified model complexity with high detection performance. This approach is suitable for object detection tasks in industrial applications with limited hardware capabilities.
[0054] In object detection, when the target is small, there is limited feature information. Using normal strided convolutions and pooling in this situation leads to the loss of granular information, a key factor contributing to the low efficiency of small-sized target detection. To address this issue, this invention introduces a convolutional method called SPD to replace the convolutional layers in the YOLOv8 network.
[0055] SPD convolution consists of a space-to-depth layer and a non-strided-convolution layer. The space-to-depth layer converts the original feature map... Perform downsampling while preserving information in the channel dimension to prevent loss of size. intermediate feature map Information.
[0056] The SPD convolutional module in S1 performs the following operations:
[0057] S11: For dimensions of Original feature map Feature sub-maps are obtained by downsampling using a scaling factor. The spatial dimension of the original feature map. The number of channels in the original feature map;
[0058] The formula for obtaining the feature sub-map in S11 is:
[0059] ;
[0060] , ;
[0061] in, For feature sub-maps, The feature map or tensor representing the input. and The subscript index of the coordinates is used to divide the sub-feature map;
[0062] S12: Connect the feature sub-maps along the channel dimension to obtain the feature map. , This is the proportionality coefficient;
[0063] S13: Add a character with The convolutional layer of the filter converts the feature map Convert to .
[0064] Each subgraph is derived from the original feature map via a scaling factor. The feature maps are obtained through downsampling. These feature sub-maps are then concatenated along the channel dimension to obtain a single feature map. The feature map's space and dimensionality are reduced by a scaling factor, while the channel dimension is increased by the square of a scaling factor. Figure 4 illustrates when... When the feature map transform value is 2, the process of transforming the feature map in the Space-to-depth layer is as follows: After the feature transformation in the Space-to-depth layer, a layer with... Filtered non-strided convolutional layers, where, The reason for using non-strided-convolution is to preserve as much discriminative information as possible. Otherwise, it would shrink the feature map, similar to using a 3×3 filter with stride=3, but each pixel would only be sampled once. If stride=2, asymmetric sampling would occur, where the sampling time for even rows / columns would be different. Generally, a stride greater than 1 leads to non-discriminative loss of information, although it may superficially appear to preserve the feature map. Convert to But no reservations were made. The characteristic information.
[0065] In the feature extraction layer, the receptive field of the network is small, resulting in weak semantic information representation ability, but strong and high-resolution representation ability for geometric details, making it suitable for perceiving location and geometric details. Conversely, deep networks have larger receptive fields and strong semantic information representation ability, but weaker geometric information representation ability and lower feature map resolution. Therefore, strengthening the exchange between high-level semantic information and low-level spatial information is key to handling different target scales. To this end, this invention proposes a novel Generalized Feature Pyramid Network (GFPN). GFPN achieves more efficient information transfer by clustering features at the same and adjacent levels, and employs skip links to prevent gradient vanishing, enhancing the ability of features to be transferred to deeper networks. While achieving a good balance between structural volume and performance, GFPN exhibits superior performance in feature fusion. The feature fusion structure diagram of GFPN is shown in Figure 5.
[0066] Because the GFPN structure has higher complexity compared to other feature fusion network structures, gradient vanishing occurs as the number of layers increases. This invention utilizes CSPStage to achieve skip connections, and simultaneously shares information between different spatial scales and non-adjacent latent semantic layers by replacing C2f (Cross Stage Partial Network Fusion) and combining it with convolutional layers. This allows the network to focus on high-level semantic information at the neck of the network while avoiding the loss of low-level spatial information.
[0067] In S2, the global feature pyramid GFPN utilizes CSPStage to achieve skip connections, while replacing C2f and combining it with convolutional layers.
[0068] The reparameterized convolution (RepConv) in the CSPStage module allows multiple computational modules to be merged during the inference phase, thereby improving the model's efficiency and performance. During training, RepConv uses convolutional layers with multiple branches. During inference, the parameters of these branches are reparameterized into the main branch, reducing computational burden and memory requirements. Therefore, implementing skip connections using CSPStage can progressively pass shallow feature information to deeper layers, minimizing feature loss and enhancing information exchange between shallow and deep layers, thus adapting to object detection at different scales.
[0069] In scenes with densely packed small targets, the original YOLOv8 detection head cannot meet the requirements for efficient and accurate detection. Therefore, this invention combines the Cross-Level Local Attention (CLLA) mechanism and Distribution Focal Loss (DFL) to design a CLLA detection head to improve the recognition and localization capabilities of targets in images.
[0070] The goal of CLLA is to model the contextual relevance of cross-level features and aggregate multi-level features, as shown in Figure 6. Features at different levels typically have different recognition information. To capture more fine-grained contextual information between different feature maps and improve the detection accuracy and robustness of the model, this invention proposes embedding a CLLA module into the detector head.
[0071] The CLLA detection head in S3 includes a CLLA cross-layer local attention mechanism layer and a DFL loss function layer, which are connected by a convolutional layer.
[0072] The CLLA cross-layer local attention mechanism layer performs the following operations:
[0073] S31: Assuming low-level feature maps and intermediate feature maps High-level feature maps contain shallow information (such as texture, edges, and color). Contains deep semantic information;
[0074] S32: Employs average pooling and Convolution reduces the size of low-level feature maps. and intermediate feature maps The spatial size is determined, and the number of channels is standardized to obtain a new low-level feature map. and new intermediate feature map Make the size and high-level feature map same;
[0075] S33: Employ learnable parameters to transform the new low-level feature map Projected to In the space, the new intermediate feature map Projected to and In the space, the formula is:
[0076] ;
[0077] ;
[0078] ;
[0079] in, , and These are query, key, and value vectors, respectively. , and This indicates that three learnable parameters are used to generate the query, key, and value vectors, respectively.
[0080] S34: Calculate using dot product and softmax function and The relevant weights, and then with Perform a dot product to obtain a new feature map;
[0081] S35: Add the new feature map to the high-level feature map Obtain aggregated feature map The formula is:
[0082] ;
[0083] in, For activation functions, superscript This is the transpose of the matrix. This is the vector dimension used to calculate the attention weights. Its purpose is to scale the dot product to prevent the dot product result from being too large, which would cause the gradient of the softmax function to become unstable.
[0084] DFL (Discrete Functionality) allows the network to quickly focus on values near the label, maximizing the probability density of the label output, guiding the model to focus on hard-to-detect targets, and improving the model's ability to detect small targets. To optimize the label... The probability of two nearby locations ( and The idea behind DFL is to use the cross-entropy function to focus the network distribution around the label value.
[0085] The DFL loss function layer uses the cross-entropy function to focus the network distribution around the label value, as shown in the formula:
[0086] ;
[0087] in, For the predicted value in and Distributed focus loss between two adjacent locations, For the network Sigmoid at position The output, For the network Sigmoid at position The output, For tags In position The probability, For tags In position The probability of.
[0088] In one embodiment of the present invention, in order to evaluate and verify the detection performance of the proposed model structure, experiments were conducted using two challenging underwater target detection datasets: the China Underwater Robot Professional Contest (URPC) dataset (2021) and the publicly available European underwater image dataset Brackish.
[0089] This invention employs precision ( Recall rate ), average accuracy ( The effectiveness of the model is evaluated using computational complexity (FLOPs), number of parameters, and frames per second (FPS). This reflects the accuracy of classifying positive samples. This reflects the effectiveness of identifying positive samples. The formula represents the average precision across all categories, FLOPs represents the number of floating-point operations that can be used to determine model complexity, and Parameters represents the number of model parameters. The formula is as follows:
[0090] ;
[0091] ;
[0092] ;
[0093] ;
[0094] in, This represents the number of positive class samples that the model correctly predicted. This represents the number of incorrectly predicted negative class samples. This represents the number of positive class samples that were incorrectly predicted. Representing the precision-recall curve, This represents the number of categories in the current identification task, and is relevant to the heavy task of object detection. Depending on the selected IoU threshold, This represents the average accuracy the model achieves on the object detection task when the IoU threshold is 0.5.
[0095] To verify the detection performance and model complexity of the proposed model and to explore the impact of specific network substructures on the model, an ablation experiment based on YOLOv8n was designed. The results are shown in Table 1, with bold text indicating the best performance.
[0096] Table 1 Ablation experimental results of URPC and Brackish datasets
[0097]
[0098] The first set of experiments primarily aimed to evaluate the detection capabilities of the original model. Subsequent individual and collective model improvement experiments, including RepGhost, SPD convolution (SPD-Conv), GFPN, and the CLLA detection head (CLLAHead), were conducted to evaluate the effectiveness of applying these four improvement techniques across two datasets. The original model achieved accuracies of 79.7% and 96.3% on the URPC and Brackish datasets, respectively. After evaluating each improvement individually, it was found that introducing the RepGhost module slightly decreased the accuracy on the URPC dataset, but effectively alleviated the problem of high model parameters and computational cost. Furthermore, RepGhost positively impacted the performance of underwater target detection in subsequent combined experiments. On the Brackish dataset, the RepGhost module positively affected both model accuracy and complexity. Additionally, GFPN showed significant performance improvements on both datasets, further demonstrating the feasibility and effectiveness of the GFPN feature fusion approach in underwater small target detection.
[0099] The proposed LFN-YOLO network achieves 82.1% and 97.4% accuracy on the URPC and Brackish datasets, respectively, representing improvements of 2.4% and 1.1% over the original model, while reducing parameter count and GFLOPs by 15.6% and 19.1%, respectively. Furthermore, the network improves recall and mAP@0.5 by 1.5% and 2.2% on the URPC dataset, and by 1.3% and 0.6% on the Brackish dataset, respectively. This demonstrates the adaptability and robustness of the LFN-YOLO model in different underwater environments.
[0100] To further illustrate that the LFN-YOLO model can better balance model complexity and accuracy, comparative experiments were conducted with eight other mainstream one-stage detection models. All experiments were performed under the same settings, using accuracy, recall, mAP@0.5, number of parameters, computational complexity, and model size to evaluate the model's performance. Specific results are shown in Table 2. All algorithms compared in the experiments meet the requirements for real-time monitoring, and LFN-YOLO demonstrates higher detection accuracy while maintaining a more lightweight model complexity.
[0101] Table 2 Comparative experiments of LFN-YOLO on URPC and Brackish datasets.
[0102]
[0103] According to the experimental results in Table 2, LFN-YOLO achieved 74.1% mAP@0.5 on the URPC dataset, which is 2.3%, 3.4%, 2.2%, 3.3%, and 8.1% higher than YOLOv5n, YOLOv6-N, YOLOv8n, YOLOv10n, and Mobilevit_v3-YOLO, respectively. Although LFN-YOLO is less lightweight than Mobilevit_v3-YOLO in terms of model size, number of parameters, and GFOLOPs, its accuracy is significantly higher. The SSD model with VGG-16 as the backbone network has the best performance on the URPC dataset, but performs poorly on the Brackish dataset. It lacks the generalization and robustness expected of an underwater target recognition model, and the SSD algorithm is relatively large in size among many one-stage algorithms, making it unsuitable for target recognition tasks on underwater unmanned platforms. It is worth noting that among the first-stage algorithms in the comparative experiments, only RetinaNet requires high hardware performance to meet the real-time requirements of underwater detection tasks; its low FPS makes it unsuitable for real-time underwater target detection. In summary, LFN-YOLO not only correctly identifies targets in underwater images but also achieves high confidence levels. It also boasts the most lightweight structure among the compared networks.
[0104] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.
Claims
1. A lightweight underwater target detection method, characterized in that, Includes the following steps: S1: Using YOLOv8 as the baseline model, the lightweight RepGhost reparameter module is used to implement feature reuse on the YOLOv8 network backbone, and the SPD convolution module is used to replace the convolutional layers of the network backbone. S2: Replace the network neck of the YOLOv8 with the global feature pyramid GFPN and perform feature fusion; S3: Replace the YOLOv8 network header with the CLLA detection head to improve target identification and localization, and obtain a lightweight underwater target detection model; S4: Evaluate the performance of the lightweight underwater target detection model and apply it to a real-world scenario to achieve underwater target detection; The CLLA detection head in S3 includes a CLLA cross-layer local attention mechanism layer and a DFL loss function layer, which are connected by a convolutional layer. The CLLA cross-layer local attention mechanism layer performs the following operations: S31: Assuming low-level feature maps and intermediate feature maps Contains shallow information, high-level feature maps Contains deep semantic information; S32: Uses average pooling Convolution reduces the size of low-level feature maps. and intermediate feature maps The spatial size is determined, and the number of channels is standardized to obtain a new low-level feature map. and new intermediate feature map Make the size and high-level feature map same; S33: Employ learnable parameters to transform the new low-level feature map Projected to In the space, the new intermediate feature map Projected to and In the space, the formula is: ； ； ； in, 、 and These are query, key, and value vectors, respectively. 、 and This indicates that three learnable parameters are used to generate the query, key, and value vectors, respectively. S34: Calculate using dot product and softmax function and The relevant weights, and then with Perform a dot product to obtain a new feature map; S35: Add the new feature map to the high-level feature map Obtain aggregated feature map The formula is: ； in, For activation functions, superscript This is the transpose of the matrix. The dimension of the vector used to compute the attention weights; The DFL loss function layer uses the cross-entropy function to focus the network distribution around the label value, as shown in the formula: ； in, For the predicted value in and Distributed focus loss between two adjacent locations, For the network Sigmoid at position The output, For the network Sigmoid at position The output, For tags In position The probability, For tags In position The probability of.
2. The lightweight underwater target detection method according to claim 1, characterized in that, The RepGhost reparameter module in S1 uses reparameterization technology to add a batch normalization layer, an Add connection layer, and an activation layer sequentially after the deformable convolutional layer, and adds a batch normalization branch between the deformable convolutional layer and the Add connection layer.
3. The lightweight underwater target detection method according to claim 1, characterized in that, The SPD convolutional module in S1 performs the following operations: S11: For dimensions of Original feature map Feature sub-maps are obtained by downsampling using a scaling factor. The spatial dimension of the original feature map. The number of channels in the original feature map; S12: Connect the feature sub-maps along the channel dimension to obtain the feature map. ， This is the proportionality coefficient; S13: Add a character with The convolutional layer of the filter converts the feature map Convert to 。 4. The lightweight underwater target detection method according to claim 3, characterized in that, The formula for obtaining the feature sub-map in S11 is: ； ， ； in, For feature sub-maps, The feature map or tensor representing the input. and This is the subscript index of the coordinate.
5. The lightweight underwater target detection method according to claim 1, characterized in that, In S2, the global feature pyramid GFPN utilizes CSPStage to achieve skip connections, while replacing C2f and combining it with convolutional layers.