An underwater target detection method based on dual-supervision guidance

Abstract

The application discloses a kind of underwater target detection methods based on double supervision guide, comprising the following steps: obtaining the data set of underwater target;Underwater target detection model based on double supervision guide is constructed, for detecting underwater target;Based on training set in data set, the double supervision training mechanism of enhancement loss and detection loss is used to the underwater target detection model, realize the close combination of image enhancement and target detection task, ensure that the image features after enhancement are more conducive to target detection, obtain the trained underwater target detection model;Test set is input into the trained underwater target detection model, and the position of underwater target in picture, target category and confidence information are output.The joint learning mechanism of double supervision guide is simultaneously optimized underwater image enhancement and target detection task, effectively improves the overall performance of underwater target detection, reduces the algorithm noise introduced in enhancement process.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of underwater target detection, and relates to an underwater target detection method based on double-supervision guidance. BACKGROUND

[0002] Underwater target detection technology has important application value in the fields of marine resource exploration, marine aquaculture monitoring, marine ecological protection, etc. However, due to the complexity of the underwater environment, light absorption and scattering cause the collected underwater images to have problems such as color distortion, low contrast, and blurred details, which seriously affect the accuracy and robustness of existing target detection algorithms. Therefore, it is a feasible method to use image enhancement to improve image quality and improve the accuracy of the target detector.

[0003] Existing underwater target detection methods mostly directly follow land target detection algorithms (such as the YOLO series and the R-CNN series), lack effective optimization measures for the degenerative features of underwater images, and the detection performance is difficult to meet the actual needs. Some existing methods that combine image enhancement and target detection only use detection loss for simple guidance, which cannot effectively improve the overall quality of the enhanced images, and may instead introduce additional noise, reducing the accuracy of target detection.

[0004] At the same time, existing target detection networks generally have problems such as large parameter quantity, high computational complexity, and poor real-time performance, making it difficult to deploy on underwater real-time application platforms with limited computing resources, such as underwater robots and remotely operated vehicles. Therefore, there is an urgent need for a target detection method that can significantly improve underwater target detection accuracy, while being lightweight and real-time, to meet the urgent needs of practical applications.

[0005] Currently, underwater target detection technology is widely used in marine resource exploration, underwater facility maintenance, and marine environmental protection. However, due to the complexity of the water environment, especially the scattering and absorption of light during underwater propagation, underwater images are severely degraded, including color deviation, image blurring, and reduced contrast. The detection accuracy and robustness of current target detection algorithms are significantly reduced when processing degraded underwater images.

[0006] In view of the deficiencies of the prior art, although existing methods attempt to improve detection performance by using underwater image enhancement (UIE) technology to improve the quality of underwater degraded images, existing methods generally have the following technical defects:

[0007] (1) Problem one: in the prior art, the underwater image enhancement task and the target detection task are independent or loosely coupled, which causes the image enhancement to be unable to specifically improve the key features required by the detection task, limiting the detection accuracy.

[0008] (2) Problem two: the existing technology generally relies only on detecting task loss for training, lacks effective supervision constraints for image enhancement tasks, and may introduce additional noise in the enhanced image, interfering with the accurate identification of the target.

[0009] (3) Problem three: the existing underwater target detection method generally has a complex model structure and large calculation amount, and is difficult to meet the needs of low power consumption and real-time in underwater practical applications. SUMMARY

[0010] To solve the above problems, the technical scheme adopted by the present application is: an underwater target detection method based on double supervision guidance, comprising the following steps:

[0011] Obtaining a data set of underwater targets;

[0012] Constructing an underwater target detection model based on double supervision guidance for detecting underwater targets;

[0013] Training the underwater target detection model based on the training set in the data set using a double supervision training mechanism of enhancement loss and detection loss, realizing the close combination of image enhancement and target detection tasks, ensuring that the features of the enhanced image are more conducive to target detection, and obtaining a trained underwater target detection model;

[0014] Inputting the test set into the trained underwater target detection model to output the position of the underwater target in the picture, the target category and the confidence information.

[0015] Further: the underwater target detection model based on double supervision guidance comprises:

[0016] A lightweight underwater image enhancement module: for adopting a double-branch structure, including a local enhancement branch and a global recovery branch; the local enhancement branch adopts a lightweight encoding-decoding structure, and uses depth separable convolution to enhance the texture and edge details of the original image; the global recovery branch uses a self-attention mechanism to recover the overall color and contrast of the original image; and the features output by the local enhancement branch and the global recovery branch are fused to obtain an enhanced image;

[0017] A double-branch backbone network: including a first branch structure and a second branch structure; the first branch structure is used to receive the original image and output feature maps of different scales; the second branch structure is used to receive the enhanced image output by the lightweight underwater image enhancement module and output feature maps of different scales;

[0018] A multi-scale attention fusion module: for adaptively fusing the feature maps of different scales output by the double-branch backbone network, using channel and spatial attention mechanisms to realize fine fusion of the feature maps;

[0019] Neck module: used for fusing different level feature maps output by the backbone network;

[0020] Head module: used for integrating the position of the underwater target in the picture, the category of the target and the confidence information based on the feature map transmitted by the neck module.

[0021] Further, the detection loss adopts the original loss function of YOLOv11, including classification loss, boundary box regression loss and distribution focus loss; the enhanced loss uses L1 loss function, taking the difference between the image and the pseudo-label clear image as the optimization goal.

[0022] Further, the local enhancement branch adopts a lightweight encoding-decoding structure, and uses depth separable convolution to enhance the texture and edge details of the original image as follows:

[0023] For a given input image I, the local branch first encodes the features through depth separable convolution to obtain shallow features Then the feature size is halved through down-sampling convolution with a step of 2, and further extracted to obtain middle-level features In the decoding stage, the features are up-sampled to the original resolution through bilinear interpolation to obtain And the shallow features in the encoding stage are connected through a jump, and the fused features are formed by concatenating in the channel dimension, and the formula is:

[0024]

[0025] Wherein: indicates concatenation in the channel dimension, and the fused features are integrated after two times of depth separable convolution, and then concatenated with the original input image to obtain Finally, the feature representation with rich local details is output through feature reorganization The formula is:

[0026]

[0027] DSC represents depth separable convolution.

[0028] Further, the global recovery branch uses a self-attention mechanism to recover the overall color and contrast of the original image as follows:

[0029] First, the input image is compressed from the original resolution to 1 / 8 of the original size through a three-level convolution down-sampling sequence, generating a compact feature representation with high semantic density

[0030] Subsequently, through the self-attention module, a global dependency model across regions is established; the attention output Z is further enriched through a linear feedforward network, and is restored to a spatial feature mapping form;

[0031] Finally, the processed features are gradually restored to the original resolution through successive three-level upsampling to obtain the final global correction features

[0032] Further, the process of fusing the features output by the local enhancement branch and the global recovery branch to obtain the enhanced image is as follows:

[0033] The feature representations of the local branch and the global branch are spliced in the channel dimension to form a composite feature representation with complete spatial-channel information. The composite feature is then subjected to cross-channel information interaction and dimension compression through a 1x1 convolution kernel to generate compact fusion features, and the formula is:

[0034]

[0035] wherein: PW represents a depth separable convolution, which has less computational complexity than ordinary convolution, models the channel relationship of the features through a channel attention mechanism, and finally maps the optimized features back to the 3-channel RGB color space through a depth separable convolution to obtain the final enhanced image

[0036] Further, the process of adaptively fusing the feature maps of different scales output by the double-branch backbone network to achieve fine fusion of the feature maps using channel and spatial attention mechanisms is as follows:

[0037] Receiving the original image features and enhanced image features extracted by the double-branch backbone network As input, first capture the statistical features of each channel through global average pooling and maximum pooling, and splice them into channel descriptors And use parallel one-dimensional convolution layers with kernel sizes of {3, 5, 7} to capture multi-scale dependency relationships within the channel. Finally, after further fusion and modeling of the nonlinear relationship through a sigmoid function, the channel attention weight W corresponding to the feature map is obtained c , the formula is:

[0038]

[0039] wherein: δ represents a sigmoid activation function, represents a one-dimensional convolution with a convolution kernel size of 1xk and a convolution kernel number of 1;

[0040] The feature map is processed in the spatial dimension to obtain the spatial attention weight W corresponding to the features ;

[0041] Channel attention weight W c Spatial attention weight W s Element-wise multiplication is performed to obtain the fusion weight W, which adaptively guides the pixel-level fusion between the two features, resulting in the fused feature representation F. fused The formula is:

[0042]

[0043] in: This indicates element-wise multiplication. Similarly, channel attention weights and spatial attention weights represent the features of the original image. This represents the channel attention weights and spatial attention weights used to enhance image features.

[0044] An underwater target detection device based on dual-supervision guidance includes:

[0045] Acquisition module: Used to acquire datasets of underwater targets;

[0046] Module: Used to build an underwater target detection model based on dual supervision, for detecting underwater targets;

[0047] Training module: The underwater target detection model is trained using a dual-supervised training mechanism of enhancement loss and detection loss based on the centralized training set of the dataset. This achieves a close integration of image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, and thus obtaining a well-trained underwater target detection model.

[0048] Detection module: This module takes the test set as input to the trained underwater target detection model and outputs the location, target category, and confidence information of the underwater target in the image.

[0049] A computer device includes: a processor and a memory, the memory storing a program module, characterized in that the program module runs on the processor to implement the method as described in any one of the claims.

[0050] An underwater target detector, characterized in that: the controller employs the method described in any one of the above methods to detect underwater targets and outputs the target's position, target category, and confidence level information in an image.

[0051] This invention provides a dual-supervised underwater target detection method. This method offers a framework for effectively integrating underwater image enhancement and underwater target detection. The underwater image enhancement module improves the quality of degraded underwater images, thereby enhancing the accuracy of underwater target detection. Simultaneously, its lightweight design results in low computational complexity, making it suitable for real-time underwater detection scenarios with limited computing resources. A dual-supervised training mechanism using enhancement and detection losses tightly integrates image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, thus effectively improving detection accuracy. By designing a dual-branch feature extraction network and introducing a multi-scale attention fusion module, feature fusion is performed on the original and enhanced images, significantly reducing the additional noise introduced by image enhancement and improving the accurate identification of targets.

[0052] By using lightweight measures such as depthwise separable convolution and simplified model structure, the number of model parameters and computational complexity are effectively reduced, meeting the requirements for low power consumption and high real-time performance in actual underwater scenarios.

[0053] Compared with the prior art, the present invention has the following significant advantages:

[0054] (1) This invention optimizes underwater image enhancement and target detection tasks simultaneously through a dual-supervised joint learning mechanism, effectively improving the overall performance of underwater target detection and reducing algorithm noise introduced in the enhancement process;

[0055] (2) The lightweight underwater image enhancement module proposed in this invention simultaneously enhances local details and global colors, thereby improving the adaptability of the enhanced image to target detection.

[0056] (3) The dual-branch backbone network and multi-scale attention fusion module designed in this invention realize the effective fusion of features of the original image and the enhanced image, further improving the detection accuracy and robustness. Attached Figure Description

[0057] 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.

[0058] Figure 1 A flowchart of an underwater target detection method based on dual supervision guidance;

[0059] Figure 2 This is a network structure diagram of the underwater target detection model DSG-Net.

[0060] Figure 3 Network structure diagram of LUIEM, a lightweight underwater image enhancement module;

[0061] Figure 4 The network structure diagram of the multi-scale attention fusion module (MSAFM);

[0062] Figure 5 A visual heatmap comparing the MSAFM multi-scale fusion module with other different fusion methods;

[0063] (a) is the input image; (b) uses the ADD method, which is to add the images directly without any fusion processing; (c) uses the Gated method, which controls the flow of information through a gating mechanism to complete the fusion; (d) uses CBAM, which is a basic method that combines channel attention and spatial attention modules for feature fusion; and (e) uses the MSAFM proposed in this application.

[0064] Figure 6 The map50-95 index shows the impact of different K values ​​of the joint loss function on the training results of the underwater target detection network DSG-Net; where (a) shows the impact of different K values ​​of the joint loss function on the training results of the underwater target detection network DSG-Net, and (b) is an enlarged view of the latter part of the curve on the left.

[0065] Figure 7 The results of the DSG-Net underwater target detection network on four publicly available experimental datasets: DUO, RUOD, UDD, and UODD. Among them, (a) the results of the experiment on the publicly available dataset DUO, (b) the results of the experiment on the publicly available dataset RUOD, (c) the results of the experiment on the publicly available dataset UDD, and (d) the results of the experiment on the publicly available dataset UODD.

[0066] Figure 8 The image shows an underwater target detector and its test results, where (a) is a schematic diagram of the detection process and (b) shows the test results of DSG-Net in various underwater scenarios. Detailed Implementation

[0067] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Figure 1 A flowchart of an underwater target detection method based on dual supervision guidance;

[0070] A dual-supervision guided underwater target detection method includes the following steps:

[0071] S1: Obtain the dataset of underwater targets;

[0072] S2: Construct an underwater target detection model based on dual supervision guidance for underwater target detection;

[0073] S3: The underwater target detection model is trained using a dual-supervised training mechanism of enhancement loss and detection loss based on the centralized training set of the dataset. This achieves a close integration of image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, and thus obtaining a well-trained underwater target detection model.

[0074] S4: Input the test set into the trained underwater target detection model and output the location of the underwater target in the image, the target category, and the confidence information.

[0075] The steps S1 / S3 / S3 / S4 are executed sequentially;

[0076] Figure 2 This is a network structure diagram of the underwater target detection model DSG-Net.

[0077] The underwater target detection model based on dual supervision includes:

[0078] Figure 3 Network structure diagram of LUIEM, a lightweight underwater image enhancement module;

[0079] Lightweight Underwater Image Enhancement Module (LUIEM): This module employs a dual-branch structure, including a local enhancement branch and a global restoration branch. The local enhancement branch uses a lightweight encoder-decoder structure and leverages depthwise separable convolution to enhance the texture and edge details of the original image. The global restoration branch uses a self-attention mechanism to restore the overall color and contrast of the original image. Finally, the enhanced image is obtained by fusing the features output from the local enhancement branch and the global restoration branch.

[0080] The dual-branch backbone network includes a first branch structure and a second branch structure. The first branch structure is used to receive the original image and output feature maps at different scales. The second branch structure is used to receive the enhanced image output by the lightweight underwater image enhancement module and output feature maps at different scales. The dual-branch backbone network uses the YOLO11 backbone structure to extract image features of two modalities respectively.

[0081] Figure 4 The network structure diagram of the multi-scale attention fusion module (MSAFM);

[0082] Multi-scale attention fusion module (MSAFM): It is used to adaptively fuse feature maps of different scales output by the dual-branch backbone network. It uses channel and spatial attention mechanisms to achieve fine fusion of feature maps, retain the effective features of the detected target, and suppress algorithm noise caused by image enhancement.

[0083] The neck module is used to fuse feature maps from different levels output by the backbone network. The input feature maps to the neck module have sizes of (64, 80, 80), (128, 40, 40), and (256, 20, 20). The neck module fuses these features, allowing for better feature representation and enhancing the model's ability to detect targets of various sizes. The output feature maps of the neck module also have sizes of (64, 80, 80), (128, 40, 40), and (256, 20, 20).

[0084] Head module: Based on the detected targets transmitted by the neck module, it outputs the underwater target's location in the image, the target's category, and confidence information. The neck module and head module belong to the YOLO11 network.

[0085] The underwater target detection model is trained end-to-end based on dual supervision. During the training process, the detection loss and enhancement loss are optimized simultaneously to guide the image enhancement module to generate image features that are more conducive to target detection.

[0086] A dual-supervised joint training network is constructed, wherein the loss function of the joint training network consists of detection loss and augmentation loss, which are jointly optimized through a dual-supervised mechanism. The detection loss adopts the original loss function of YOLOv11, including classification loss, bounding box regression loss and distribution focus loss. The augmentation loss uses the L1 loss function, with the optimization objective being to enhance the difference between the image and the clear image with pseudo-labels.

[0087] The trained model is deployed to a computationally limited underwater edge computing device, where real-time underwater images are input into the network for real-time detection.

[0088] The local enhancement branch employs a lightweight encoder-decoder structure, and the process of enhancing the texture and edge details of the original image using depthwise separable convolution is as follows:

[0089] For a given input image I, the local branch first performs feature encoding through depthwise separable convolution to obtain shallow features that preserve the original resolution. The feature size is then halved using a downsampling convolution with a stride of 2, and mid-layer features are further extracted. During the decoding stage, features are upsampled to the original resolution using bilinear interpolation. and shallow features in the encoding stage Perform skip connections and concatenate them along the channel dimension to form a fused feature, as shown in the formula:

[0090]

[0091] in This indicates that the concatenation is performed along the channel dimension. The fused features are integrated through two depthwise separable convolutions and then concatenated with the original input image to obtain the final image. Finally, feature reshaping outputs a feature representation with rich local details. The formula is:

[0092]

[0093] DSC stands for Depthwise Separable Convolution;

[0094] The global restoration branch uses a self-attention mechanism to restore the overall color and contrast of the original image as follows:

[0095] First, the input image is progressively compressed from its original resolution to 1 / 8 of its original size using a three-level convolutional downsampling sequence, generating a compact feature representation with high semantic density.

[0096] Subsequently, a global dependency model across regions is established through a self-attention module; the attention output Z is then further enriched through a linear feedforward network and restored to a spatial feature mapping form.

[0097] Finally, the processed features are gradually restored to the original resolution through three consecutive levels of upsampling, resulting in the final globally corrected features.

[0098] The process of fusing the features output by the local enhancement branch and the global recovery branch to obtain the enhanced image is as follows:

[0099] The feature representations of the local and global branches are concatenated along the channel dimension to form a composite feature representation with complete spatial-channel information. This composite feature is then subjected to cross-channel information interaction and dimensionality compression through a 1×1 convolution kernel to generate a compact fused feature, as shown in the formula:

[0100]

[0101] Where: PW represents depthwise separable convolution, which has less computational cost than ordinary convolution. It models the channel relationships of features through a channel attention mechanism, and finally maps the optimized features back to the 3-channel RGB color space through depthwise separable convolution to obtain the final enhanced image.

[0102] The process of adaptively fusing feature maps of different scales output by the dual-branch backbone network, and utilizing channel and spatial attention mechanisms to achieve fine-grained feature map fusion, is as follows:

[0103] Receive the original image features and enhanced image features extracted by the dual-branch backbone network. As input, the statistical features of each channel are first captured by global average pooling and max pooling respectively, and then concatenated into channel descriptors. The parallel one-dimensional convolutional layers with kernel sizes {3,5,7} are used to capture multi-scale dependencies within channels. Finally, after further fusion and modeling nonlinear relationships using the sigmoid function, the channel attention weights W corresponding to the feature maps are obtained. c The formula is expressed as:

[0104]

[0105] Where δ represents the sigmoid activation function. This represents a one-dimensional convolution with a kernel size of 1×k and a kernel size of 1.

[0106] The feature map is processed in the spatial dimension to obtain the spatial attention weights W corresponding to the features. s ;

[0107] Channel attention weight W c Spatial attention weight W s Element-wise multiplication is performed to obtain the fusion weight w, which adaptively guides the pixel-level fusion between the two features, resulting in the fused feature representation F. fused The formula is:

[0108]

[0109] in: This indicates element-wise multiplication. Similarly, channel attention weights and spatial attention weights represent the features of the original image. This represents the channel attention weights and spatial attention weights used to enhance image features.

[0110] An underwater target detection device based on dual-supervision guidance includes:

[0111] Acquisition module: Used to acquire datasets of underwater targets;

[0112] Module: Used to build an underwater target detection model based on dual supervision, for detecting underwater targets;

[0113] Training module: The underwater target detection model is trained using a dual-supervised training mechanism of enhancement loss and detection loss based on the centralized training set of the dataset. This achieves a close integration of image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, and thus obtaining a well-trained underwater target detection model.

[0114] Detection module: This module takes the test set as input to the trained underwater target detection model and outputs the location, target category, and confidence information of the underwater target in the image.

[0115] A computer device includes: a processor and a memory, the memory storing a program module, characterized in that the program module runs on the processor to implement the method as described in any one of the claims.

[0116] An underwater target detector, characterized in that: the controller uses any one of the methods described above to detect underwater targets and outputs the target's position, target category, and confidence level information in the image.

[0117] Example 1: In the actual implementation of this invention, the publicly available datasets DUO, RUOD, UDD, and UODD were used to conduct experiments on the proposed dual-supervised guided network (DSG-Net). Ablation comparison experiments were performed on the image enhancement module and the multi-scale attention fusion module, respectively, and performance comparison experiments were conducted with current mainstream object detection methods to demonstrate the effectiveness and superiority of the technical solution of this invention. The specific implementation steps are as follows:

[0118] S1: Dataset Preparation: Specific implementation experiments were conducted using publicly available underwater target detection datasets, such as DUO. The image categories in this dataset cover multiple typical underwater targets such as sea cucumbers, sea urchins, scallops, and starfish. The dataset was divided into training, validation, and test sets in a ratio of 7:2:1 for network training and performance validation.

[0119] S2: Construct an underwater target detection model based on dual supervision guidance for underwater target detection;

[0120] The underwater target detection model includes a lightweight image enhancement module, a dual-branch backbone network, a multi-scale attention fusion module, a neck module, and a head module. The specific construction process is as follows:

[0121] Construction of the image enhancement module: A lightweight underwater image enhancement module (LUIEM) is constructed, which adopts a dual-branch structure. The local enhancement branch uses a lightweight improved structure of U-Net, utilizing depthwise separable convolutions to enhance local image details; the global recovery branch uses a self-attention module to capture the overall color and contrast features of the image. The two branches are finally fused to output the enhanced image. In this embodiment, the output image size of the image enhancement module is uniformly 640×640, which serves as the input image for the dual-branch network.

[0122] Construction of the dual-branch backbone network: Using YOLOv11's backbone as a base, it is extended into a dual-branch backbone network. The first branch processes the original image, and the second branch processes the enhanced image. The enhanced image provides a clearer feature representation, while the original image compensates for the negative impact of texture destruction in the enhanced image. The final output is feature maps of different scales from both branches.

[0123] Construction of the multi-scale attention fusion module: The features output by the two branches in the dual-branch backbone network are adaptively fused by the multi-scale attention fusion module (MSAFM) designed in this invention. This module utilizes channel and spatial attention mechanisms to achieve fine fusion of feature maps, thereby improving the feature differences between the detected target and the background.

[0124] The fused feature maps are input into the YOLO11 Neck and Header network structures. After post-processing, the final localization information of the target in the image is obtained.

[0125] S3: Dual-Supervised Joint Training Mechanism: During training, this invention jointly optimizes the underwater target detection network parameters through a joint constraint loss consisting of detection loss and enhancement loss. The detection loss uses the original loss function of YOLOv11, including classification loss, bounding box regression loss, and focus distribution loss. The enhancement loss uses the L1 loss function, with the optimization objective being to enhance the difference between the enhanced image and the clear image with pseudo-labels. In the experiment, by adjusting the weight coefficient k of the enhancement loss, the optimal value was determined to be 5, under which the network achieved its best performance.

[0126] During training, the network was trained using an NVIDIA GeForce RTX 3090 GPU for a total of 600 epochs with an initial learning rate of 0.01. Random flipping and Mosaic data augmentation techniques were used to improve the model's generalization ability. After training, the performance metrics of the detection model on the test set were evaluated using the standard COCO evaluation method. Finally, the obtained model was converted into a Tensorrt-supported model and deployed to edge computing devices commonly used in underwater robots such as the Jetson AGX Xavier.

[0127] S4: Input the test set into the trained underwater target detection model and output the location, target category, and confidence information of the underwater target in the image.

[0128] This invention proposes a dual-supervised guided network model (DSG-Net) for underwater target detection, which can improve the detection performance of small-sized underwater targets while maintaining the detection effect of normal-sized targets, and significantly reduce the number of model parameters and computational complexity. Figure 2As shown, this invention first proposes a lightweight underwater image enhancement module (LUIEM), comprising a local enhancement branch and a global restoration branch, to effectively restore local details and global color features of underwater images, generating high-quality images more suitable for target detection. Secondly, a dual-branch backbone network structure is designed to independently extract features from the original and enhanced images, achieving collaborative expression of features from different modalities. Then, a multi-scale attention fusion module (MSAFM) is proposed to replace traditional feature fusion methods, enabling full interaction and fusion between high-level semantic information and low-level spatial information, minimizing algorithmic noise interference introduced during enhancement. Finally, a joint loss function (JC-Loss) based on a dual-supervision mechanism is designed, combining detection loss and enhancement loss to effectively improve the network's target recognition accuracy and localization capability, ultimately forming a dual-supervision guided underwater target detection model (DSG-Net) suitable for resource-constrained edge computing platforms.

[0129] The LUIEM module mentioned in the dual-supervised guided network model receives the original image. As input, the output is the enhanced image. In the algorithm, H and W are generally taken as fixed dimensions of 640, and the input and output dimensions are the same. Its network structure is as follows: Figure 3 As shown.

[0130] Step 1-1: Input the original image The inputs are fed into the local branch to enhance local details such as texture, and into the global branch to enhance global features such as contrast and color cast.

[0131] Steps 1-2: For a given input image I, the local branch first performs feature encoding through depthwise separable convolution to obtain shallow features that preserve the original resolution. The feature size is then halved using a downsampling convolution with a stride of 2, and mid-layer features are further extracted. During the decoding stage, features are upsampled to the original resolution using bilinear interpolation. and shallow features in the encoding stage Perform skip connections and concatenate them along the channel dimension to form a fused feature, as shown in the formula:

[0132]

[0133] in: This indicates that the concatenation is performed along the channel dimension. The fused features are integrated through two depthwise separable convolutions and then concatenated with the original input image to obtain the final image. Finally, feature reshaping outputs a feature representation with rich local details. The formula is:

[0134]

[0135] Steps 1-3: For the global branch, the input image is first compressed from its original resolution to 1 / 8 of its original size using a three-level convolutional downsampling sequence, generating a compact feature representation with high semantic density. Subsequently, a cross-regional global dependency model is established using a self-attention module. The attention output Z is then further enriched through a linear feedforward network and restored to a spatial feature map form. Finally, the processed features are progressively restored to the original resolution through three consecutive levels of upsampling, yielding the final globally corrected features.

[0136] Steps 1-4: After obtaining the feature representations of the local and global branches, the two are concatenated along the channel dimension to form a composite feature representation with complete spatial-channel information. This composite feature is then processed by a 1×1 convolution kernel for cross-channel information interaction and dimensionality compression, generating a compact fused feature. The formula is as follows:

[0137]

[0138] Here, PW represents depthwise separable convolution, which has less computational cost than ordinary convolution and is one of our lightweight methods. Then, a channel attention mechanism is used to model the channel relationships of the features, and finally, depthwise separable convolution maps the optimized features back to the 3-channel RGB color space to obtain the final enhanced image.

[0139] The network structure of the MSAFM module is as follows: Figure 4 As shown, it performs the following operations:

[0140] Step 3-1: Accept the original image features and enhanced image features extracted by the dual backbone network. As input, the statistical features of each channel are first captured by global average pooling and max pooling respectively, and then concatenated into channel descriptors. The parallel one-dimensional convolutional layers with kernel sizes {3,5,7} are used to capture multi-scale dependencies within channels. Finally, after further fusion and modeling nonlinear relationships using the sigmoid function, the channel attention weights W corresponding to the feature maps are obtained. c The formula is expressed as:

[0141]

[0142] Where: δ represents the sigmoid activation function. This represents a one-dimensional convolution with a kernel size of 1×k and a kernel count of 1.

[0143] Step 3-2: Similarly, process the feature map in the spatial dimension to obtain the spatial attention weights W corresponding to the features. s .

[0144] Step 3-3: Adjust the channel attention weights W c Spatial attention weight W s Element-wise multiplication is performed to obtain the fusion weight w, which adaptively guides the pixel-level fusion between the two features, resulting in the fused feature representation F. fused The formula is:

[0145]

[0146] in This indicates element-wise multiplication.

[0147] We compared MSAFM with several other fusion methods and visualized the fusion results using Grad-CAM++. Figure 5 As shown, (a) is the input image; (b) uses the ADD method, which involves direct addition without any fusion processing; (c) uses the Gated method, which controls the flow of information through a gating mechanism to complete the fusion; (d) uses CBAM, a basic method that combines channel attention and spatial attention modules for feature fusion; and (e) uses our proposed MSAFM, which adds multi-scale feature perception to CBAM to enhance the fusion capability of the module. It can be seen that our method achieves better image fusion results and more clearly identifies the regions of interest of the target object.

[0148] Dual-supervised joint learning is the core of the method. It guides the UIE module to generate enhanced features more suitable for the detection task through Joint-Constraint Loss (JC-Loss), thus achieving global feature enhancement. JC-Loss consists of detection loss and enhancement loss. The detection loss uses the YOLOv11 detection loss function, with the following formula:

[0149]

[0150] in: Indicates the bounding box loss. Represents classification loss, This represents Distribution FocalLoss, used to address the class imbalance problem in object detection. To reduce the impact of noise in false labels, we employ a more robust method. The loss function measures the difference between the augmented image and the pseudo-label, and the formula is:

[0151]

[0152] Among them, I enhanced Indicates an enhanced image, I target Let N represent the number of pixel values, where N represents the number of sharp image pseudo-labels. The JC-Loss can be obtained by combining the proportions of the detection loss and the enhancement loss, and can be expressed as:

[0153]

[0154] Where k is the dynamic coefficient. Because our training process requires clear labels for the underwater degraded images in the object detection dataset, which is difficult to achieve in practice, we utilize existing underwater image enhancement models to generate clear image pseudo-labels for the underwater degraded images as supervision information. It helps the UIE module obtain effective feature information from clear image pseudo-labels, but it also introduces noise interference from the pseudo-labels. The UIE module is constrained by the detection performance to retain more detection-advantageous features and reduce noise interference in pseudo-labels.

[0155] The effect of different k values ​​on the training results is shown in the curve. Figure 6 As shown, (a) shows the impact of different K values ​​of the joint loss function on the training results of the underwater target detection network DSG-Net, quantified using the map50-95 index; (b) is an enlarged view of the latter part of the curve on the left. It can be seen that when the joint loss function is used, i.e., K≠0, the training effect is always better than when K=0, i.e., the map50-95 value is higher when K≠0, indicating higher recognition accuracy.

[0156] We use map50-95 to measure the model's detection accuracy; the higher the value, the higher the detection accuracy. As we can see, when using the joint loss function (k≠0), the training performance is always better than when using only the detection loss (k=0). This proves that our proposed dual-supervised guided training method outperforms traditional single-supervised training methods.

[0157] We conducted related experiments on the DUO dataset, and all metrics were calculated using the COCO metric. The red and blue numbers represent the best and second-best performance for the current metric, respectively.

[0158] Experiment 1, LUIEM Ablation Experiment: To verify the effectiveness of our proposed UIE module LUIEM, we investigated different UIE methods. We set up five sets of experiments, where the original image was input to branch 1 of the network backbone and the enhanced image was input to branch 2. The experiments compared two enhancement strategies: a physical model-based UIE method and a learning-based UIE method. The learning-based method was jointly learned with the detector. In addition, a set of experiments using the original image as input was set up as a benchmark experiment. For the physical model-based method, we used CBF, an underwater image enhancement method based on color balance and multi-scale fusion. By compensating for the attenuation of the red and blue channels and combining white balance and edge sharpening, it effectively improves the contrast, color, and detail clarity of underwater images. The learning-based UIE method used FA+Net and LiteEnhanceNet, which are lightweight underwater image enhancement networks. Our proposed LUIEM also belongs to the learning-based method. The experimental results are shown in the table below:

[0159]

[0160] Experiment 1, serving as the foundational experiment, achieved an AP of 64.6 without any enhancement methods. Experiment 2, using CBF as an enhancement method, improved the AP by 0.1%. Experiments 3, 4, and 5 further improved performance using learning-based enhancement methods: FA+Net improved the AP by 0.5%, LiteEnhanceNet by 0.6%, and our proposed ULEM achieved the best performance, improving the AP of the foundational method by 1.0%.

[0161] Experiment 2, MSAFM Ablation Experiment: To verify the effectiveness of MSAFM, we set up four experiments, each using a different fusion method to fuse the feature maps of the two main branches. Experiment 1 (ADD) served as the baseline experiment and did not use any fusion method, directly adding the feature maps of the two main branches. Experiment 2 (Gated) employed a gate-based fusion strategy. Experiment 3 (CBAM) used an attention-based fusion method. Experiment 4 (MSAFM) used our proposed multi-scale attention fusion method. The experimental results are shown in the table below:

[0162]

[0163] Experimental results show that the addition of a fusion strategy is beneficial to improving detector performance. The Gated method improved the model's AP by 0.3%, while using only channel and spatial attention mechanisms improved the model's AP by 0.2%. Our proposed MSAFM method enhances the basic dual attention mechanism with multi-scale perception, improving the model's AP by 0.6%, achieving the best results.

[0164] Experiment 3, Baseline Performance Comparison Experiment: The accuracy performance of our proposed DSG-Net on the DUO dataset is compared with other commonly used models in the following table:

[0165]

[0166] As we can see, our DSG-Net outperforms other models overall, exhibiting better detection accuracy.

[0167] Our visualization detection results on the four datasets DUO, RUOD, UDD, and UODD are as follows: Figure 7 As shown, (a) the experimental results of the proposed dual-supervised underwater target detection model on the publicly available dataset DUO, (b) the experimental results of the proposed dual-supervised underwater target detection model on the publicly available dataset RUOD, (c) the experimental results of the proposed dual-supervised underwater target detection model on the publicly available dataset UDD, and (d) the experimental results of the proposed dual-supervised underwater target detection model on the publicly available dataset UODD.

[0168] As you can see, our proposed DSG-Net can effectively detect underwater targets. We also conducted robot entity tests in real-world underwater scenarios, such as... Figure 8 As shown, (a) is a schematic diagram of the detection process, and (b) shows the test results of DSG-Net in various underwater scenarios. It can be seen that even in complex and varied real-world underwater scenarios, our proposed method can accurately locate the target.

[0169] The proposed dual-supervision guided underwater target detection method effectively improves the accuracy and robustness of underwater target detection by introducing a dual supervision mechanism of detection loss and enhancement loss. Experimental results show that on the publicly available DUO dataset, the average accuracy (AP) of this method reaches 65.6%, compared to YOLOv11's 63.7%, representing a 1.9% improvement. On the RUOD dataset, the AP reaches 61.8%, a 2.4% improvement over the baseline method. On the UDD and UODD datasets, the APs reach 31.2% and 51.8%, respectively, representing improvements of 3.0% and 2.8%. Compared to other mainstream general-purpose detectors, it also shows accuracy improvements ranging from 2% to 20%. These results demonstrate that this invention can significantly improve the performance indicators of target detection in underwater environments, clearly outperforming existing mainstream detection algorithms.

[0170] Furthermore, this invention employs a lightweight network architecture design, with only 3.9M parameters and a computational cost of only 15.9 GFLOPs. Compared to existing typical underwater target detection methods such as GCC-Net and general detectors like DETR, the number of parameters and computational cost are reduced by more than 90%. On a Jetson AGX Xavier edge computing device, this invention can achieve a real-time inference speed of approximately 40 FPS, significantly outperforming existing methods and greatly improving the efficiency and real-time performance of underwater target detection tasks.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 underwater target detection method based on dual-supervision guidance, characterized in that: Includes the following steps: Obtain a dataset of underwater targets; An underwater target detection model based on dual supervision is constructed for detecting underwater targets; The underwater target detection model is trained using a dual-supervised training mechanism of enhancement loss and detection loss based on the centralized training set of the dataset. This achieves a close integration of image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, and thus obtaining a well-trained underwater target detection model. The test set is input into the trained underwater target detection model, which outputs the location of the underwater target in the image, the target category, and the confidence information. The underwater target detection model based on dual supervision includes: Lightweight underwater image enhancement module: It employs a dual-branch structure, including a local enhancement branch and a global restoration branch. The local enhancement branch uses a lightweight encoder-decoder structure and leverages depthwise separable convolution to enhance the texture and edge details of the original image. The global restoration branch uses a self-attention mechanism to restore the overall color and contrast of the original image. The enhanced image is obtained by fusing the features output from the local enhancement branch and the global restoration branch. The dual-branch backbone network includes a first branch structure and a second branch structure. The first branch structure is used to receive the original image and output feature maps at different scales. The second branch structure is used to receive the enhanced image output by the lightweight underwater image enhancement module and output feature maps at different scales. Multi-scale attention fusion module: used to adaptively fuse feature maps of different scales output by the dual-branch backbone network, and to achieve fine fusion of feature maps by utilizing channel and spatial attention mechanisms; Neck module: Used to fuse feature maps from different levels output by the backbone network; Head module: Used to integrate the feature maps transmitted by the neck module and output the underwater target's location in the image, target category, and confidence information.

2. The underwater target detection method based on dual-supervision guidance according to claim 1, characterized in that: The detection loss uses the original loss function of YOLOv11, including classification loss, bounding box regression loss, and distribution focus loss; the enhancement loss uses the L1 loss function, with the optimization objective being to enhance the difference between the image and the clear image with pseudo-labels.

3. The underwater target detection method based on dual-supervision guidance according to claim 1, characterized in that: The local enhancement branch employs a lightweight encoder-decoder structure, and the process of enhancing the texture and edge details of the original image using depthwise separable convolution is as follows: For a given input image The local branches first encode features using depthwise separable convolutions to obtain shallow features that maintain the original resolution. The feature size is then halved using a downsampling convolution with a stride of 2, and the mid-layer features are further extracted. During the decoding stage, features are upsampled to the original resolution using bilinear interpolation. and shallow features in the encoding stage A skip connection is made, and the features are spliced ​​together along the channel dimension to form a fused feature. The formula is: in: This indicates that the concatenation is performed along the channel dimension. The fused features are integrated through two depthwise separable convolutions and then concatenated with the original input image to obtain the final image. Finally, feature reshaping outputs a feature representation with rich local details. The formula is: DSC stands for Depthwise Separable Convolution.

4. The underwater target detection method based on dual-supervision guidance according to claim 1, characterized in that: The global restoration branch uses a self-attention mechanism to restore the overall color and contrast of the original image as follows: First, the input image is progressively compressed from its original resolution to 1 / 8 of its original size using a three-level convolutional downsampling sequence, generating a compact feature representation with high semantic density. ; Subsequently, a global dependency model across regions is established through a self-attention module; the attention output Z is then further enriched in feature representation through a linear feedforward network and restored to a spatial feature mapping form. Finally, the processed features are gradually restored to the original resolution through three consecutive levels of upsampling, resulting in the final globally corrected features. .

5. The underwater target detection method based on dual-supervision guidance according to claim 1, characterized in that: The process of fusing the features output by the local enhancement branch and the global recovery branch to obtain the enhanced image is as follows: The feature representations of the local and global branches are concatenated along the channel dimension to form a composite feature representation with complete spatial-channel information. This composite feature is then passed through a 1×1 convolutional kernel for cross-channel information interaction and dimensionality compression, generating a compact fused feature, as shown in the formula: Where: PW represents depthwise separable convolution, which has less computational cost than ordinary convolution. It models the channel relationships of features through a channel attention mechanism, and finally maps the optimized features back to the 3-channel RGB color space through depthwise separable convolution to obtain the final enhanced image. .

6. The underwater target detection method based on dual-supervision guidance according to claim 1, characterized in that: The process of adaptively fusing feature maps of different scales output by the dual-branch backbone network, and utilizing channel and spatial attention mechanisms to achieve fine-grained feature map fusion, is as follows: Receive the original image features and enhanced image features extracted by the dual-branch backbone network. As input, the statistical features of each channel are first captured by global average pooling and max pooling respectively, and then concatenated into channel descriptors. It utilizes parallel one-dimensional convolutional layers with kernel sizes {3,5,7} to capture multi-scale dependencies within channels, and finally, after further fusion and... After modeling the nonlinear relationship using the function, the channel attention weights corresponding to the feature map are obtained. The formula is expressed as: in: express Activation function Indicates the kernel size as A one-dimensional convolution with a kernel size of 1; The feature map is processed in the spatial dimension to obtain the spatial attention weights corresponding to the features. ; Channel attention weights Spatial attention weights Perform element-wise multiplication to obtain the fusion weights. This adaptively guides pixel-level fusion between two features, resulting in a fused feature representation. The formula is: in: This indicates element-wise multiplication. Similarly, channel attention weights and spatial attention weights represent the features of the original image. This represents the channel attention weights and spatial attention weights used to enhance image features.

7. An underwater target detection device based on dual-supervision guidance, characterized in that: include: Acquisition module: Used to acquire datasets of underwater targets; Module: Used to build an underwater target detection model based on dual supervision, for detecting underwater targets; The underwater target detection model based on dual supervision includes: Lightweight underwater image enhancement module: It employs a dual-branch structure, including a local enhancement branch and a global restoration branch. The local enhancement branch uses a lightweight encoder-decoder structure and leverages depthwise separable convolution to enhance the texture and edge details of the original image. The global restoration branch uses a self-attention mechanism to restore the overall color and contrast of the original image. The enhanced image is obtained by fusing the features output from the local enhancement branch and the global restoration branch. The dual-branch backbone network includes a first branch structure and a second branch structure. The first branch structure is used to receive the original image and output feature maps at different scales. The second branch structure is used to receive the enhanced image output by the lightweight underwater image enhancement module and output feature maps at different scales. Multi-scale attention fusion module: used to adaptively fuse feature maps of different scales output by the dual-branch backbone network, and to achieve fine fusion of feature maps by utilizing channel and spatial attention mechanisms; Neck module: Used to fuse feature maps from different levels output by the backbone network; Head module: Used to integrate the feature maps transmitted by the neck module and output the underwater target's location in the image, target category, and confidence information; Training module: The underwater target detection model is trained using a dual-supervised training mechanism of enhancement loss and detection loss based on the centralized training set of the dataset. This achieves a close integration of image enhancement and target detection tasks, ensuring that the enhanced image features are more conducive to target detection, and thus obtaining a well-trained underwater target detection model. Detection module: This module takes the test set as input to the trained underwater target detection model and outputs the location, target category, and confidence information of the underwater target in the image.

8. A computer device, comprising: A processor and a memory, the memory storing a program module, characterized in that the program module runs on the processor to implement claim 1. The method described in any one of the 6 methods.

Images