AUV underwater target detection method based on c2f_smpcglu lightweight network
By constructing the C2f_SMPCGLU lightweight network, the problems of accuracy, real-time performance, and lightweight deployment in underwater target detection are solved, achieving high-precision detection of underwater targets with weak texture and small scale, and adapting to the limited computing resources on the AUV side.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV CHINA
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies cannot simultaneously meet the requirements of target detection accuracy in complex underwater environments and the needs of real-time inference and lightweight deployment at the AUV end-side. In particular, they are not effective in detecting underwater targets with weak textures and small scales, and the number of model parameters and computational load are large.
A lightweight C2f_SMPCGLU network is adopted. By constructing a lightweight SMPCGLU feature extraction unit and embedding it into a C2f structure to form a C2f_SMPCGLU module, a lightweight backbone network is constructed by combining a layer-by-layer decreasing large kernel convolution strategy, a multi-scale feature fusion detection head and RTDETRDecoder decoding to achieve multi-scale feature extraction and target classification.
It achieves high-precision detection of weakly textured and small-scale targets in complex underwater environments, while significantly reducing the number of model parameters and computational complexity, meeting the needs of real-time inference and lightweight deployment on the AUV edge.
Smart Images
Figure CN122391818A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater target detection technology, and in particular to an AUV underwater target detection method based on a lightweight C2f_SMPCGLU network. Background Technology
[0002] Autonomous Underwater Vehicles (AUVs), with their autonomous operating capabilities, are widely used in marine resource exploration, underwater emergency search and rescue, seabed facility inspection, and intelligent underwater operations. Underwater visual perception technology provides crucial data support for AUVs' environmental recognition, autonomous navigation, and intelligent decision-making, becoming a core technology for ensuring the efficiency and safety of underwater operations. Underwater imaging environments are affected by factors such as light attenuation, color distortion, scattering from suspended particles, and background noise. Images generally exhibit characteristics such as low contrast, blurred edges, difficulty in effectively representing target features, and small scale, placing higher demands on the accuracy and adaptability of underwater target detection technology.
[0003] Current underwater target detection methods mostly rely on general deep learning convolutional neural network frameworks, directly transferring land scene detection models to underwater scenes, completing feature extraction through conventional convolution operators, performing feature representation based on fixed-structure feature modules, processing feature maps at different levels using general multi-scale feature extraction methods, and improving the accuracy of detection output with complex network structures.
[0004] The fixed sampling method of general detection models is difficult to adapt to the edge and local texture variations of underwater targets with weak texture. The feature module has limited enhancement effect on underwater target features and cannot fully suppress background redundancy. The multi-scale feature extraction stage lacks adaptive receptive field design, making it difficult to stably identify underwater targets of different scales. These models usually have a large number of parameters and computational load, and when deployed on AUV end-user platforms, they suffer from poor computing power adaptability and limited inference speed, failing to simultaneously meet the usage requirements of target detection accuracy in complex underwater environments and real-time inference and lightweight deployment on end-user platforms. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides an AUV underwater target detection method based on a lightweight C2f_SMPCGLU network, which solves the problem that existing technologies cannot simultaneously meet the requirements of target detection accuracy in complex underwater environments and real-time inference and lightweight deployment at the edge.
[0006] This invention provides an AUV underwater target detection method based on a lightweight C2f_SMPCGLU network, including an offline model training stage and an end-side real-time detection stage;
[0007] The offline model training phase includes: Acquire an underwater target image, and preprocess the underwater target image to obtain a preprocessed underwater target image; The preprocessed underwater target image is input into a lightweight backbone network to obtain multiple feature maps of different scales; wherein, the lightweight backbone network is constructed with the C2f_SMPCGLU module as the core unit, and the C2f_SMPCGLU module is formed by embedding the SMPCGLU lightweight feature extraction unit into the C2f structure; Multiple feature maps of different scales are input into the detection head for feature fusion to obtain enhanced multi-scale features; The multi-scale features are input into RTDETRDecoder for target classification and bounding box regression to construct a lightweight target detection model. The end-side real-time detection phase includes: Acquire real-time underwater images, and preprocess the real-time underwater images to obtain preprocessed real-time underwater images; The preprocessed underwater real-time image is input into the lightweight detection model, and the target detection result is output.
[0008] Furthermore, the preprocessing of the underwater target image includes: The underwater target image is subjected to size scaling, pixel normalization, and information annotation processing.
[0009] Furthermore, the processing steps of the SMPCGLU lightweight feature extraction unit include: The input features are fed into the SMPConv convolutional layer for convolution operations and activation function processing to obtain intermediate features; The intermediate features are input into a convolutional gated linear unit for feature enhancement processing to obtain enhanced features; The enhanced features are processed by random depth regularization and then added to the input features by residual addition to obtain the output features of the SMPCGLU lightweight feature extraction unit.
[0010] Furthermore, the C2f_SMPCGLU module is formed by embedding the SMPCGLU lightweight feature extraction unit into the C2f structure, including: The standard basic units within the branches of the C2f structure are replaced with SMPCGLU lightweight feature extraction units to form the C2f_SMPCGLU module.
[0011] Furthermore, the feature processing steps of the C2f_SMPCGLU module include: The input features are fed into the channel branch for processing, resulting in retained branch features and initial features; The initial features are input into the SMPCGLU lightweight feature extraction unit to obtain the enhanced initial features; The preserved branch features and the enhanced initial features are fused to obtain fused features. The fused features are then convolved to obtain the module output features.
[0012] Furthermore, the construction steps of the lightweight backbone network include: Replace multiple C2f modules at different levels in the YOLOv8 backbone network with multiple C2f_SMPCGLU modules.
[0013] Furthermore, the feature processing steps of the lightweight backbone network include: The preprocessed underwater target image is subjected to convolutional downsampling to obtain shallow features; By setting different numbers of C2f_SMPCGLU modules at different levels of the lightweight backbone network to enhance the features of the previous layer input, the enhanced features are obtained. The last layer of deep features is input into the SPPF module to improve the multi-scale expressive power of deep features; The lightweight backbone network outputs multiple feature maps of different scales.
[0014] Furthermore, the feature maps at multiple different scales include shallow feature maps, mid-level feature maps, and deep feature maps; The step of the detection head performing feature fusion on multiple feature maps of different scales includes: The deep feature map is upsampled by a preset factor to obtain an upsampled deep feature map. The resolution of the upsampled deep feature map is the same as that of the mid-layer feature map. The upsampled deep feature map and the mid-layer feature map are then fused by channel splicing to obtain an enhanced mid-layer fused feature map. The mid-layer fusion feature is upsampled by a preset factor to obtain an upsampled mid-layer feature map. The upsampled mid-layer feature map is then fused with the shallow feature map through channel splicing to obtain an enhanced shallow fusion feature. Enhanced deep fusion features are obtained based on deep feature maps; The output is enhanced multi-scale features, which include shallow fusion features, mid-level fusion features and deep fusion features.
[0015] Furthermore, after inputting the multi-scale features into RTDETRDecoder for target classification and bounding box regression, a model network structure is obtained. The preprocessed underwater target image is then input into the model network structure, and the network parameters are iteratively optimized using a detection loss function to construct a lightweight target detection model.
[0016] Furthermore, the detection results include category, bounding box location, and confidence level.
[0017] Compared with the prior art, the present invention has the following advantages: This application constructs a lightweight feature extraction unit, SMPCGLU, and embeds it into a C2f structure to form a C2f_SMPCGLU module. A lightweight backbone network is built around this module. By combining a layer-by-layer decreasing large kernel convolution strategy, a multi-scale feature fusion detection head, and RTDETRDecoder decoding, high-precision detection of targets with weak texture, small scale, and low contrast in complex underwater environments is achieved. At the same time, the number of model parameters and computational complexity are significantly reduced, meeting the requirements of real-time inference, lightweight deployment, and low-power operation on the AUV edge. This effectively solves the problem that existing technologies cannot balance detection accuracy and edge-side practical performance.
[0018] Based on the above reasons, this invention can be widely applied in fields such as underwater target detection. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a diagram illustrating the overall framework of the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of this invention. Figure 2 This is a structural diagram of the SMPCGLU module in the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of the present invention; Figure 3 This is a diagram showing the internal structure of the C2f_SMPCGLU network in the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of this invention. Figure 4 This is a diagram of the lightweight backbone network structure of the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of the present invention. Detailed Implementation
[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0022] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.
[0023] Autonomous Underwater Vehicles (AUVs), with their autonomous operating capabilities, are widely used in marine resource exploration, underwater emergency search and rescue, seabed facility inspection, and intelligent underwater operations. Underwater visual perception technology provides crucial data support for AUVs' environmental recognition, autonomous navigation, and intelligent decision-making, becoming a core technology for ensuring the efficiency and safety of underwater operations. Underwater imaging environments are affected by factors such as light attenuation, color distortion, scattering from suspended particles, and background noise. Images generally exhibit characteristics such as low contrast, blurred edges, difficulty in effectively representing target features, and small scale, placing higher demands on the accuracy and adaptability of underwater target detection technology.
[0024] Current underwater target detection methods mostly rely on general deep learning convolutional neural network frameworks, directly transferring land scene detection models to underwater scenes, completing feature extraction through conventional convolution operators, performing feature representation based on fixed-structure feature modules, processing feature maps at different levels using general multi-scale feature extraction methods, and improving the accuracy of detection output with complex network structures.
[0025] The fixed sampling method of general detection models is difficult to adapt to the edge and local texture variations of underwater targets with weak texture. The feature module has limited enhancement effect on underwater target features and cannot fully suppress background redundancy. The multi-scale feature extraction stage lacks adaptive receptive field design, making it difficult to stably identify underwater targets of different scales. These models usually have a large number of parameters and computational load, and when deployed on AUV end-user platforms, they suffer from poor computing power adaptability and limited inference speed, failing to simultaneously meet the usage requirements of target detection accuracy in complex underwater environments and real-time inference and lightweight deployment on end-user platforms.
[0026] This invention provides an AUV underwater target detection method based on the C2f_SMPCGLU lightweight network, which solves the problem that existing technologies cannot simultaneously meet the requirements of target detection accuracy in complex underwater environments and real-time inference and lightweight deployment at the edge.
[0027] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0028] Please see Figure 1 , Figure 1 This is a diagram illustrating the overall framework of the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of this invention.
[0029] This application provides an AUV underwater target detection method based on the C2f_SMPCGLU lightweight network, including an offline model training stage and an edge-side real-time detection stage.
[0030] The offline model training phase includes the following steps: Step 101: Acquire underwater target images and preprocess them to obtain preprocessed underwater target images. This process corrects the problems of light attenuation, color distortion, and noise interference from suspended particles caused by water absorption and scattering. It also enhances the edge details and feature representation capabilities of underwater targets, providing high-quality input for feature extraction in the subsequent lightweight backbone network. This improves the robustness of the detection model to complex underwater environments and the detection accuracy of small-scale, weakly textured targets, ensuring the stability of model training and the reliability of real-time detection at the edge.
[0031] In some embodiments, the underwater target image is preprocessed, including: scaling the underwater target image, pixel normalization, and information annotation. Scaling unifies the image input size, eliminating interference from size differences on feature extraction and ensuring consistency in model training; pixel normalization maps pixel values to a uniform range, accelerating model gradient convergence and improving network training stability; information annotation provides accurate labels for supervised model training, ensuring the model effectively learns underwater target features and improving the accuracy and robustness of subsequent target detection.
[0032] Specifically, let the original input image be... The preprocessed image is denoted as Then we have:
[0033] in, This indicates an image resizing operation. This indicates a normalization operation.
[0034] Step 102: Input the preprocessed underwater target image into the lightweight backbone network to obtain multiple feature maps of different scales; the lightweight backbone network is constructed with the C2f_SMPCGLU module as the core unit, and the C2f_SMPCGLU module is formed by embedding the SMPCGLU lightweight feature extraction unit into the C2f structure.
[0035] This structure utilizes the gating mechanism of SMPCGLU units to dynamically filter feature information of underwater targets with weak texture, and combines it with a C2f multi-branch structure to preserve the original features. This significantly reduces the number of model parameters and computational complexity while achieving multi-level feature enhancement. The lightweight backbone network constructed in this way can efficiently extract multi-scale features of underwater targets and enhance the feature representation capability of small-scale, weak-textured targets.
[0036] like Figure 2 As shown, Figure 2 This is a structural diagram of the SMPCGLU module in the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of the present invention.
[0037] In some embodiments, the processing steps of the SMPCGLU lightweight feature extraction unit include: inputting the input features into an SMPConv convolutional layer for convolution and activation function processing to obtain intermediate features; inputting the intermediate features into a convolutional gated linear unit for feature enhancement processing to obtain enhanced features; and adding the enhanced features to the input features after random depth regularization to obtain the output features of the SMPCGLU lightweight feature extraction unit.
[0038] The specific process is as follows: First, the input feature x is input into the SMPConv convolutional layer, and after processing by the activation function, intermediate features are obtained. , can be represented as:
[0039] in, Act represents the convolution operation. This indicates the activation function processing operation. SMPConv convolutional layers can expand the feature receptive field with extremely low computational overhead, enhance the ability to extract local texture and edge details of weakly textured underwater targets, and provide high-quality intermediate features for subsequent feature enhancement.
[0040] Then the intermediate features Input convolutional gated linear units to perform feature filtering and enhancement, resulting in enhanced features. , can be represented as:
[0041] in, This represents a convolutional gated linear unit operation. ConvolutionalGLU dynamically selects effective features through a gating mechanism, suppressing redundant background noise interference such as underwater suspended particles, and significantly improving the feature representation capability of underwater targets.
[0042] Finally, the enhanced features will be implemented. After random deep layer regularization, and with the input features Perform residual summation to obtain the module output characteristics. , can be represented as:
[0043] in, This indicates a random depth regularization operation. DropPath random depth can effectively alleviate the overfitting problem during the training of deep networks and improve the generalization ability of the model; residual connections ensure the stable propagation of gradients in the network, avoid gradient vanishing, and significantly improve the stability and convergence efficiency of model training.
[0044] In some embodiments, the C2f_SMPCGLU module is formed by embedding SMPCGLU lightweight feature extraction units into a C2f structure, including replacing the standard basic units within the branches of the C2f structure with SMPCGLU lightweight feature extraction units to form the C2f_SMPCGLU module. This replacement structure can enhance the feature representation capability of underwater weakly textured targets through the gating mechanism of SMPCGLU units while retaining the advantages of C2f's multi-branch feature extraction, and significantly reduce the number of module parameters and computational complexity. The C2f_SMPCGLU module constructed in this way can effectively adapt to the limited computing resources on the AUV side while achieving feature enhancement at different layers, meeting the real-time detection requirements, and providing core support for feature extraction of the subsequent lightweight backbone network.
[0045] like Figure 3 As shown, Figure 3 This is a diagram of the internal structure of the C2f_SMPCGLU network in the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of this invention.
[0046] In some embodiments, the feature processing steps of the C2f_SMPCGLU module include: inputting the input features into a channel branch for processing to obtain retained branch features and initial features; inputting the initial features into the SMPCGLU lightweight feature extraction unit to obtain enhanced initial features; fusing the retained branch features and enhanced initial features to obtain fused features; and performing a convolution operation on the fused features to obtain the module output features.
[0047] The specific process is as follows: Input features are First, the channel branching process is performed to obtain the branch-preserving feature. With initial features , can be represented as:
[0048] in, This represents the channel branching operation in the C2f structure, which is used to split the input features into two parallel branches, preserving the original feature information while providing input for feature enhancement.
[0049] Then, multiple SMPCGLU modules are cascaded in the enhancement branch to perform step-by-step feature enhancement on the initial features. The output of each SMPCGLU module is denoted as , can be represented as:
[0050] in, SMPCGLU represents the lightweight feature extraction unit operation, which is used to perform lightweight convolution extraction and gating enhancement on input features. It enhances the feature representation of underwater weak textures and small-scale targets while reducing the amount of computation, and ensures stable gradient propagation through residual structure.
[0051] After completing multi-level feature enhancement, branch features will be retained. Output characteristics at each level of the enhanced branch The channels are then fused using a fusion process followed by convolution to obtain the output features of the C2f_SMPCGLU module. , can be represented as:
[0052] in, This represents a multi-branch feature fusion operation, used to fuse original features with multi-level enhanced features to enrich feature dimensions and information hierarchy; This indicates the fused convolutional operation, used to unify feature dimensions, optimize feature representation, and ensure consistency in feature extraction from the subsequent backbone network; The number of times the SMPCGLU modules are stacked can be adaptively adjusted according to the computing power requirements of the AUV end, achieving a balance between detection accuracy and inference speed.
[0053] In some embodiments, the construction steps of the lightweight backbone network include replacing multiple C2f modules at different levels in the YOLOv8 backbone network with multiple C2f_SMPCGLU modules. While retaining the original network's feature extraction capabilities, this significantly reduces the number of parameters and computational complexity of the backbone network, improves the feature representation capabilities for underwater targets with weak textures and small scales, makes the model more adaptable to the limited computing power of AUV end-users, and meets the requirements for real-time detection and lightweight deployment.
[0054] like Figure 4 As shown, Figure 4 This is a diagram of the lightweight backbone network structure of the AUV underwater target detection method based on the C2f_SMPCGLU lightweight network of the present invention.
[0055] In some embodiments, the feature processing steps of the lightweight backbone network include: performing convolutional downsampling on the preprocessed underwater target image to obtain shallow features; enhancing the features of the previous layer input by setting different numbers of C2f_SMPCGLU modules at different levels of the lightweight backbone network to obtain enhanced features; inputting the last layer of deep features into the SPPF module to improve the multi-scale expressive power of deep features; and outputting multiple feature maps of different scales from the lightweight backbone network.
[0056] Specifically, to balance receptive field modeling capability and computational efficiency, this invention adopts a layer-by-layer decreasing large kernel convolution strategy. That is, as the network layer deepens, the module convolution kernel size gradually decreases from 13 to 7, while the number of feature channels gradually increases. This achieves collaborative optimization of receptive field extraction of global context and small convolution kernel enhancement of local details at different layers, effectively adapting to the multi-scale feature extraction needs of underwater targets.
[0057] In the four downsampling stages of P2 / 4, P3 / 8, P4 / 16 and P5 / 32, a corresponding number of C2f_SMPCGLU modules are configured to achieve accurate extraction and enhancement of underwater target features at different scales. The SPPF module is connected to the end of the backbone network, and deep features are aggregated through spatial pyramid pooling to further enhance the multi-scale expression capability of deep features and provide high-quality input for feature fusion of subsequent detection heads.
[0058] In this embodiment, after feature extraction at different levels, the feature maps output by the backbone network are denoted as follows: The hierarchical relationship can be represented as follows:
[0059]
[0060]
[0061]
[0062] in, The input image after preprocessing. Indicates the first The process involves downsampling and feature enhancement stages. Finally, the backbone network outputs data to the detection head. The three scale feature maps are used for subsequent multi-scale feature fusion, target classification and bounding box regression, providing multi-scale and high-discrimination feature support for underwater target detection at the AUV end.
[0063] Step 103: Input feature maps of multiple different scales into the detection head for feature fusion to obtain enhanced multi-scale features. By fusing and complementing information across different scales, the richness and robustness of feature representation are effectively improved, enhancing the distinguishability of feature representation for small-scale, low-contrast underwater targets, and providing high-precision and high-reliability feature support for subsequent target classification and bounding box regression.
[0064] In some embodiments, the multiple feature maps of different scales include shallow feature maps, mid-level feature maps, and deep feature maps; wherein... It is a shallow feature map, containing rich details such as target edges and textures; It is a mid-level feature map that takes into account both details and semantic information; It is a deep feature map, which contains strong semantic information but relatively little detailed information.
[0065] The steps of feature fusion for multiple feature maps at different scales by the detection head include: upsampling the deep feature map by a preset factor to obtain an upsampled deep feature map, the resolution of which is consistent with that of the mid-level feature map; fusing the upsampled deep feature map and the mid-level feature map through channel concatenation to obtain an enhanced mid-level fused feature; upsampling the mid-level fused feature map by a preset factor to obtain an upsampled mid-level feature map; fusing the upsampled mid-level feature map and the shallow feature map through channel concatenation to obtain an enhanced shallow fused feature; obtaining the enhanced deep fused feature based on the deep feature map; and outputting the enhanced multi-scale features, which include shallow fused features, mid-level fused features, and deep fused features.
[0066] The specific fusion process is as follows: First, the deep feature map is processed... Perform upsampling at a preset factor, which can be 2x, to make its resolution comparable to that of the mid-level feature map. The two are matched and fused through channel concatenation to obtain the enhanced intermediate fused feature. , can be represented as:
[0067] This step, through the complementary fusion of deep semantic features and mid-level detailed features, not only preserves the global contextual information of deep features but also supplements the target details of mid-level features, effectively improving the feature discrimination ability of mid-level features for underwater targets.
[0068] Then the intermediate fusion features Perform upsampling again by 2 times, and compare with the shallow feature map. Channel splicing and fusion are performed to obtain enhanced shallow fusion features. :
[0069] in, For a 2x upsampling operation, This is for channel splicing operations.
[0070] This operation further integrates multi-level enhancement features and shallow detail features, enhancing the edge details and feature representation of small-scale, weakly textured underwater targets.
[0071] To maintain multi-scale detection capability, the detection head ultimately outputs enhanced multi-scale features. ,in It can be derived from deep features You can keep it directly.
[0072] After completing the multi-scale feature fusion, the enhanced multi-scale features will be... The data is fed into the RTDETRDecoder for end-to-end decoding, and the target category prediction result of the decoded output is denoted as... The bounding box regression result is denoted as Then we have:
[0073] in, This indicates the target category prediction result. This represents the bounding box position regression result. This decoding process, with lightweight computational overhead, enables accurate identification and localization of multi-scale targets in complex underwater scenarios, significantly improving the accuracy, robustness, and real-time performance of underwater target detection on the AUV end-side, laying a core foundation for subsequent end-side deployment and practical applications.
[0074] Step 104: Input multi-scale features into RTDETRDecoder for target classification and bounding box regression to construct a lightweight target detection model. This decoding process efficiently integrates multi-scale feature information through the Transformer global attention mechanism, significantly reducing computation time while improving the accuracy of underwater target classification and the precision of bounding box regression, thus enhancing the model's robustness and generalization ability in complex underwater environments. The lightweight detection model constructed in this way can achieve efficient inference under limited computing resources on the AUV end, meeting real-time detection requirements, and ultimately achieving high-precision and high-reliability detection of multi-scale, weakly textured underwater targets, thus fulfilling the technical objectives of this invention.
[0075] In some embodiments, multi-scale features are input into the RTDETRDecoder for target classification and bounding box regression to obtain the model network structure. Preprocessed underwater target images are then input into the model network structure, and the network parameters are iteratively optimized using a detection loss function to construct a lightweight target detection model. This optimization process significantly reduces the number of network parameters and computational complexity while ensuring model detection accuracy, accelerating gradient convergence during model training, and improving the stability and generalization ability of network parameter optimization. The lightweight detection model constructed in this way can effectively adapt to the limited computing resources on the AUV end-user and meet real-time detection requirements.
[0076] The edge-side real-time detection phase includes: Step 201: Acquire underwater real-time images and preprocess them to obtain preprocessed underwater real-time images.
[0077] During missions, the AUV uses its onboard underwater vision sensor to acquire images of the surrounding environment in real time. The raw underwater images are first preprocessed, including scaling and pixel normalization, to standardize the size and numerical range required by the lightweight detection network. This eliminates interference from image quality degradation and size differences, ensuring consistency in subsequent feature extraction. The preprocessed images are then input into the offline-trained lightweight detection network. The backbone network performs multi-scale feature extraction, and the detection head performs upsampling, channel stitching, and multi-level feature fusion. Finally, the RTDETRDecoder outputs the target category, bounding box location, and corresponding confidence score.
[0078] Let the lightweight detection network that has been trained offline and deployed on the AUV end be denoted as . The real-time detection output can then be expressed as:
[0079] in, Indicates time Preprocessed images, Indicates time Target category prediction results Indicates time The location of the target bounding box. Indicates time The corresponding detection confidence level.
[0080] Because this invention employs a collaborative design of the C2f_SMPCGLU lightweight module and progressively decreasing large kernel convolution, it effectively reduces the number of model parameters and computational complexity while ensuring the ability to extract multi-scale features from underwater targets. This achieves a balance between detection accuracy and inference efficiency, thereby meeting the engineering requirements of real-time inference, lightweight deployment, and low-power operation on the AUV end-side. It provides highly reliable and real-time visual detection support for environmental perception in AUV autonomous underwater operations.
[0081] Step 202: Input the preprocessed real-time underwater image into the lightweight detection model and output the target detection result. In some embodiments, the detection result includes category, bounding box location, and confidence score.
[0082] After completing network inference, the system outputs the target location, category, and confidence level, achieving real-time underwater target detection. The detection results can then be sent to the AUV's upper-level mission system for target perception, environmental analysis, path planning, and autonomous operation decision-making.
[0083] This application constructs a lightweight feature extraction unit, SMPCGLU, and embeds it into a C2f structure to form a C2f_SMPCGLU module. A lightweight backbone network is built around this module. By combining a layer-by-layer decreasing large kernel convolution strategy, a multi-scale feature fusion detection head, and RTDETRDecoder decoding, high-precision detection of targets with weak texture, small scale, and low contrast in complex underwater environments is achieved. At the same time, the number of model parameters and computational complexity are significantly reduced, meeting the requirements of real-time inference, lightweight deployment, and low-power operation on the AUV edge. This effectively solves the problem that existing technologies cannot balance detection accuracy and edge-side practical performance.
[0084] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0085] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An AUV underwater target detection method based on a lightweight C2f_SMPCGLU network, characterized in that, This includes the offline model training phase and the edge-side real-time detection phase; The offline model training phase includes: Acquire an underwater target image, and preprocess the underwater target image to obtain a preprocessed underwater target image; The preprocessed underwater target image is input into a lightweight backbone network to obtain multiple feature maps of different scales; wherein, the lightweight backbone network is constructed with the C2f_SMPCGLU module as the core unit, and the C2f_SMPCGLU module is formed by embedding the SMPCGLU lightweight feature extraction unit into the C2f structure; Multiple feature maps of different scales are input into the detection head for feature fusion to obtain enhanced multi-scale features; The multi-scale features are input into RTDETRDecoder for target classification and bounding box regression to construct a lightweight target detection model. The end-side real-time detection phase includes: Acquire real-time underwater images, and preprocess the real-time underwater images to obtain preprocessed real-time underwater images; The preprocessed underwater real-time image is input into the lightweight detection model, and the target detection result is output.
2. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, The preprocessing of the underwater target image includes: The underwater target image is subjected to size scaling, pixel normalization, and information annotation processing.
3. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, The processing steps of the SMPCGLU lightweight feature extraction unit include: The input features are fed into the SMPConv convolutional layer for convolution operations and activation function processing to obtain intermediate features; The intermediate features are input into a convolutional gated linear unit for feature enhancement processing to obtain enhanced features; The enhanced features are processed by random depth regularization and then added to the input features by residual addition to obtain the output features of the SMPCGLU lightweight feature extraction unit.
4. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, The C2f_SMPCGLU module is formed by embedding the SMPCGLU lightweight feature extraction unit into the C2f structure, including: The standard basic units within the branches of the C2f structure are replaced with SMPCGLU lightweight feature extraction units to form the C2f_SMPCGLU module.
5. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 4, characterized in that, The feature processing steps of the C2f_SMPCGLU module include: The input features are fed into the channel branch for processing, resulting in retained branch features and initial features; The initial features are input into the SMPCGLU lightweight feature extraction unit to obtain the enhanced initial features; The preserved branch features and the enhanced initial features are fused to obtain fused features. The fused features are then convolved to obtain the module output features.
6. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, The construction steps of the lightweight backbone network include: Replace multiple C2f modules at different levels in the YOLOv8 backbone network with multiple C2f_SMPCGLU modules.
7. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 6, characterized in that, The feature processing steps of the lightweight backbone network include: The preprocessed underwater target image is subjected to convolutional downsampling to obtain shallow features; By setting different numbers of C2f_SMPCGLU modules at different levels of the lightweight backbone network to enhance the features of the previous layer input, the enhanced features are obtained. The last layer of deep features is input into the SPPF module to improve the multi-scale expressive power of deep features; The lightweight backbone network outputs multiple feature maps of different scales.
8. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, Multiple feature maps at different scales include shallow feature maps, medium feature maps, and deep feature maps; The step of the detection head performing feature fusion on multiple feature maps of different scales includes: The deep feature map is upsampled by a preset factor to obtain an upsampled deep feature map. The resolution of the upsampled deep feature map is the same as that of the mid-layer feature map. The upsampled deep feature map and the mid-layer feature map are then fused by channel splicing to obtain an enhanced mid-layer fused feature map. The mid-layer fusion feature is upsampled by a preset factor to obtain an upsampled mid-layer feature map. The upsampled mid-layer feature map is then fused with the shallow feature map through channel splicing to obtain an enhanced shallow fusion feature. Enhanced deep fusion features are obtained based on deep feature maps; The output is enhanced multi-scale features, which include shallow fusion features, mid-level fusion features and deep fusion features.
9. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, After inputting the multi-scale features into RTDETRDecoder for target classification and bounding box regression, the model network structure is obtained. The preprocessed underwater target image is then input into the model network structure, and the network parameters are iteratively optimized using the detection loss function to construct a lightweight target detection model.
10. The AUV underwater target detection method based on the C2f_SMPCGLU lightweight network according to claim 1, characterized in that, The detection results include category, bounding box location, and confidence level.