Underwater noise submarine cable detection method and related device

Abstract

The invention discloses an underwater noise submarine cable detection method and a related device. The method comprises the following steps: acquiring an underwater target image, and constructing a data set; the training set is input into a BBT-RTDETR model for training, a trained BBT-RTDETR model is obtained, and a backbone network in the BBT-RTDETR model is used for performing feature extraction on the input image by adopting multilayer convolution and a CAMDT module to obtain multi-scale features; the encoder is used for performing layer-by-layer feature extraction and scale conversion through the RepC3 residual module, the BBT module and the up-down sampling module, and performing multi-scale semantic fusion on the multi-scale feature map in cooperation with multiple jump connections to obtain multi-scale semantic fusion features; the decoder and the prediction head are used for target detection and classification based on the multi-scale semantic fusion features. According to the invention, the submarine cable detection precision in an underwater noise environment can be improved.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method and related apparatus for detecting underwater noise in submarine cables. Background Technology

[0002] Underwater target detection technology is one of the core technologies in the fields of marine engineering and resource development, with applications covering submarine cable inspection, pipeline maintenance, and marine biological monitoring. Due to the complex optical scattering, low contrast, and high noise interference (such as suspended sediment, biological disturbance, and equipment noise) in the underwater environment, traditional computer vision-based methods (such as edge detection and threshold segmentation) often suffer from significantly reduced detection accuracy due to insufficient feature extraction capabilities. In recent years, deep learning-based target detection models (such as the YOLO series (You Only Look Once) and Faster R-CNN (Faster Region-based Convolutional Neural Network)) have improved detection performance through end-to-end feature learning; however, image degradation caused by underwater noise (such as Gaussian noise, salt-and-pepper noise, and motion blur) still leads to a sharp decrease in the model's generalization ability.

[0003] To address the noise robustness issue, existing research mainly employs data augmentation (such as noise injection, GAN (Generative Adversarial Network) training) or adaptive methods, but these suffer from problems such as large deviations between simulated noise and the real environment, and redundant model structures. Transformer architectures, due to their global attention mechanism, demonstrate advantages in complex scene modeling. For example, the DETR (Detection Transformer) series of models achieves end-to-end detection through an encoder-decoder structure, but its computational complexity is high and it is sensitive to local features. While RTDETR, as a real-time detection model, optimizes inference speed, it still faces challenges such as insufficient feature interaction capabilities and a lack of noise suppression mechanisms under underwater noise interference. Furthermore, Bayesian deep learning methods enhance the model's uncertainty perception ability by introducing probabilistic modeling, and image reconstruction techniques combined with Gaussian Process Regression (GPR) can effectively compensate for semantic information in noisy regions; however, existing research has not fully integrated these methods into lightweight detection frameworks. How to construct a detection system that balances real-time performance, noise robustness, and embedded deployment requirements has become a technical bottleneck for intelligent underwater cable detection. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an underwater noise submarine cable detection method and related apparatus, which enables submarine cable detection to perform well even under underwater noise conditions, thereby improving the accuracy of submarine cable detection in underwater noise environments.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for detecting underwater noise in submarine cables includes the following steps:

[0007] Acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio;

[0008] The training set is input into a pre-defined BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features.

[0009] The underwater target image to be detected is input into the trained BBT-RTDETR model to obtain the target detection and classification results.

[0010] Preferably, the processing procedure of the CAMDT module includes:

[0011] The input features are extracted using a series of sequentially connected AMDT modules based on a fusion channel attention mechanism and a spatial attention mechanism to obtain depth transformation features.

[0012] The depth transformation features are concatenated with the input features to obtain the fused features;

[0013] Channel compression is performed on the fused features to obtain the output features of the CAMDT module.

[0014] Preferably, the AMDT module includes an AMCT module, a layer normalization module, a dilated convolution module, a GeLU activation layer, a dilated convolution module, and an AMST module connected in sequence. The output of the AMCT module is residually concatenated with the output of the AMST module to obtain the output of the AMDT module.

[0015] The processing procedure of the AMCT module includes: the original input X aAfter layer normalization, the input is fed into three parallel convolutional layers. The outputs of two of these convolutional layers are concatenated to obtain the first channel feature. This first channel feature is then processed by the SE attention module to obtain a channel attention weight vector. This channel attention weight vector is then multiplied channel-by-channel with the output of the third convolutional layer to obtain the feature vector. ; the features Compared with the original input X a By splicing

[0016] The processing procedure of the AMST module includes: the original input X b After layer normalization, the input is fed into three parallel dilated convolutional layers, each using a different dilation rate. The outputs of two of these layers are concatenated to obtain a first spatial feature. This first spatial feature is then concatenated with the output of the third dilated convolutional layer to obtain a multi-scale spatial feature. This multi-scale spatial feature is then processed using max pooling and average pooling to obtain two spatial feature maps. These two feature maps are concatenated and processed through a convolutional layer to generate a spatial attention mask. This spatial attention mask is then multiplied pixel-by-pixel with the output of the third dilated convolutional layer to obtain the feature map. ; the features Compared with the original input X b By splicing .

[0017] Preferably, the training process is divided into four training stages. In the first training stage, the classification branch parameters are frozen, and training is performed only by minimizing the localization loss to optimize the localization branch parameters. In the second training stage, the localization branch parameters are frozen, and training is performed only by minimizing the classification loss to optimize the classification branch parameters. In the third training stage, the parameter freezing is lifted, and the model is trained using joint loss. In the fourth training stage, Bayesian optimization is performed by introducing ELBO loss on the basis of joint loss, and Gaussian noise data augmentation is added to train the model.

[0018] Preferably, the localization loss includes GIoU loss and L1 loss, the classification loss uses Varifocal Loss loss, and the joint loss includes GIoU loss, L1 loss, and Varifocal Loss loss.

[0019] Preferably, the method further includes the following steps:

[0020] Determine the evaluation metrics and evaluate the performance of the BBT-RTDETR model using a test set;

[0021] The evaluation metric is the mean accuracy (mAP).

[0022] Preferably, Gaussian noise of different intensities is added to the images in the test set.

[0023] Based on the above, the present invention also discloses an underwater noise submarine cable detection system, comprising:

[0024] The acquisition module is used to acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio.

[0025] The training module is used to input the training set into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features.

[0026] The detection module is used to input the underwater target image to be detected into the trained BBT-RTDETR model to obtain the target detection and classification results.

[0027] Based on the foregoing, the present invention also discloses a computer device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement any of the methods described above.

[0028] Based on the above, the present invention also discloses a readable storage medium storing a computer program, which, when executed by a processor, implements any of the methods described above.

[0029] Based on the above technical solution, the beneficial effects of the present invention are as follows: The present invention discloses an underwater noise submarine cable detection method, which acquires underwater target images, constructs a dataset, and divides the dataset into a training set and a test set according to a preset ratio; the training set is input into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features; the encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module, and simultaneously performs multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network using multiple skip connections to obtain multi-scale semantic fusion features; the decoder and prediction head are used to perform target detection and classification based on the multi-scale semantic fusion features; the underwater target image to be detected is input into the trained BBT-RTDETR model to obtain target detection and classification results. This invention enables submarine cable detection to perform well even under underwater noise conditions through a multi-scale attention mechanism and a Bayesian optimization model, thereby improving the accuracy of submarine cable detection in underwater noise environments. Attached Figure Description

[0030] Figure 1 This is a schematic flowchart of an underwater noise submarine cable detection method in one embodiment;

[0031] Figure 2 This is a schematic diagram of the structure of the BBT-RTDETR model in one embodiment;

[0032] Figure 3 Here is a schematic diagram of the CAMDT module in one embodiment, wherein (a) is a schematic diagram of the AMCT module; (b) is a schematic diagram of the AMST module; (c) is a schematic diagram of the AMDT module; and (d) is a schematic diagram of the CAMDT module.

[0033] Figure 4 This is a schematic diagram of the structure of the BBT module in one embodiment. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0035] like Figure 1 As shown in the figure, this embodiment provides a method for detecting underwater noise in submarine cables, including the following steps:

[0036] Step 100: Acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio.

[0037] In this embodiment, 2000 underwater cable images were collected to construct a dataset, of which 1600 images were used for training and 400 images were used for testing. To simulate a real underwater environment, Gaussian noise of varying intensities was added to the images in the test set.

[0038] Step 200: Input the training set into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features.

[0039] In this embodiment, the pre-defined BBT-RTDETR (Bayesian-Enhanced Multi-attention Deconvolutional Transformer Network) model comprises three parts: Backbone, Encoder, and Decoder & Head. See also... Figure 2Solid arrows represent the feature flow process along different paths. The backbone, encoder, decoder, and prediction head are identified with blue, red, and green wireframes, respectively. The backbone performs initial feature extraction on the input image, which is a 640×640×3 RGB image. After processing through multiple convolutional layers and CAMDT (Cross-Stage Part-Attention Multi-scale Dual Transformer) modules, multi-scale feature maps of 320×320×64, 160×160×128, 80×80×256, 40×40×384, and 20×20×384 are extracted. The four orange CAMDT modules introduce a contextual attention mechanism to enhance the expressive power between multi-scale features. The final output features are compressed into a 20×20×256 feature map via BBT and convolution before being fed into the encoder. The encoder section is used to construct deep semantic features. First, the input features are processed by convolution and a Bayesian Boosted Transformer (BBT) module. Then, layer-by-layer feature extraction and scale transformation are performed through a RepC3 (Reparameterized C3) residual module and upsampling / downsampling modules. Simultaneously, multiple skip connections pass feature maps of different resolutions to the final fusion module, achieving multi-scale semantic fusion. The decoder and prediction head section receives features from multiple layers and concatenates them before feeding them into the decoder and head section for final prediction, suitable for downstream tasks such as detection or segmentation. This network structure effectively combines attention mechanisms, multi-scale feature fusion, and multi-path design, significantly improving the model's ability to express image details and global semantic information. A detailed description of the structure and function of each part of the BBT-RTDETR model is as follows (see [link]). Figure 3 .

[0040] 1) AMCT module (Attention Multiscale Channel Converter)

[0041] The AMCT (Attention Multi-scale Channel Transformer) module takes input data from the feature map of the previous layer, with dimensions H×W×C (height×width×number of channels). In terms of processing, the input feature map X is first processed by parallel 3×3 convolutional kernels. aThe input features are processed to extract channel features at different granularities. Each convolutional branch independently processes the input features, resulting in multiple intermediate feature representations. Then, the output features from the two right-hand sides are concatenated or added along the channel dimension to form a fused multi-scale channel feature representation. This is reflected in Figure 3 (a) The upper right fusion region, with dimensions H×W×C'. Then, using the SE (Squeeze-and-Excitation) module, the feature map is first compressed using Global Average Pooling (GAP). The vector, embodied in Figure 3 (a) The blue matrix part, then the dependencies between channels are learned through two fully connected layers to generate channel attention weight vectors. (size is) Finally, the channel attention weight vector The output of the third convolutional layer is multiplied channel by channel, which is reflected in... Figure 3 (a) The lower left fusion region enhances important channels and suppresses redundant channels to obtain output features. ; final characteristics Compared with the original input X a By splicing .

[0042] The processing result is the output feature map after channel attention calibration. The size is The feature responses of key channels are enhanced. The feature maps output by the AMCT module will be passed to the AMST (Attention Multi-Scale Spatial Transformer) module or directly participate in the subsequent feature fusion process, providing channel-selective high-quality feature representations for spatial attention processing and multi-scale feature integration, and enhancing the model's ability to capture information at different semantic levels.

[0043] 2) AMST module (Attention Multiscale Spatial Transformer)

[0044] The AMST (Attention Multi-scale Spatial Transformer) module takes feature maps from the AMCT module or the previous layer as input, with dimensions H×W×C. In terms of processing, it first employs multiple parallel dilated convolution branches, each using a different dilation rate (e.g., dilation rate = 1, 2, 3), thereby capturing spatial contextual information within different receptive fields without increasing the number of parameters. Each branch outputs an intermediate feature representation. Then, the outputs of the two dilated convolution branches on the right are concatenated or added along the channel dimension to form a fused multi-scale spatial feature. This is reflected in Figure 3 (b) The upper right fusion area, with dimensions of Next, by analyzing the fusion features... Perform max pooling and average pooling operations along the channel dimension to obtain two... The spatial feature maps are then concatenated and passed through a convolutional layer to generate a spatial attention mask. (Dimensions are H×W×1), reflected in Figure 3 (b) The blue matrix portion, this mask represents the importance of each spatial location in the feature map. Finally, the spatial attention mask is applied. The output of the third dilated convolutional layer on the left is multiplied pixel-by-pixel, which is reflected in... Figure 3 (b) The lower left fusion region enhances the features of important spatial areas and suppresses the response of background and noise areas to obtain the output features. ;at last Concatenate with the original input to obtain .

[0045] The processing result is the output feature map after spatial attention calibration. The size is The feature responses of the target region and key spatial locations are enhanced. The feature map output by the AMST module is passed to the AMDT (Attention Multi-Scale Dual Transformer) module for deep fusion with channel attention features, or directly participates in the multi-scale feature integration of the feature encoder, providing spatially selective fine feature representations for subsequent target detection and localization.

[0046] 3) AMDT module (Attention Multiscale Dual Transformer)

[0047] The input data for the AMDT (Attention Multi-scale Dual Transformer) module is the feature map from the previous layer, with a size of [missing information]. It may have already undergone preliminary processing by AMCT or AMST. In terms of processing methods, the AMDT module includes, in sequence, an AMCT module, layer normalization, dilated convolution, GeLU activation layer, dilated convolution, and an AMST module (such as...). Figure 3 (as shown in (c)). In terms of processing, the input features first undergo channel attention enhancement via the AMCT module to obtain... Then, feature transformation is performed sequentially through layer normalization, 3×3 dilated convolution, GeLU activation layer, and 3×3 dilated convolution; finally, spatial attention enhancement is performed through the AMST module to obtain... By adding the original input feature map to the AMST output through residual connections, the original semantic information is preserved and information loss is prevented, resulting in the final output. .

[0048] The processing result is a feature map that has undergone dual channel and spatial attention enhancement. The size is It possesses both channel selectivity and spatial selectivity, enabling more accurate focusing on target regions and key features. The feature maps output by the AMDT module are passed to the CAMDT (Cross-Stage Part-Attention Multi-Scale Dual Transformer) module for cross-stage feature splitting and fusion, or directly participate in the deep feature extraction process of the encoder, providing the model with multi-scale dual-enhanced features with strong discriminative capabilities, significantly improving the modeling ability for complex underwater noise scenes.

[0049] 4) CAMDT Module (Cross-Stage Part - Attention Multi-Scale Dual Transformer)

[0050] The CAMDT (Cross-Stage Partial) module takes input data from feature maps of size H×W×C from the layer preceding the Backbone or Encoder. In terms of processing, it first employs a CSP (Cross-Stage Partial) structure, dividing the input feature map into two parts along the channel dimension. One part (usually half a channel) is directly passed to the output via a skip connection, while the other part enters the AMDT module for complex multi-scale dual attention transformation. This design reduces computational redundancy while preserving some original feature information. The feature map entering the AMDT branch undergoes the aforementioned parallel processing of channel attention and spatial attention, along with dual feature enhancement operations, generating depth-transformed features. The features then transformed by AMDT will be... The original features passed with jump connections The data is stitched together along the channel dimension to form a fused feature. The size is (Or adjust according to the diversion ratio). Finally, through Convolution performs channel compression on the concatenated features, reducing the number of channels to the target dimension. The final output features are obtained. The size is .

[0051] The processing result is a feature map that has undergone cross-stage feature fusion and channel compression. The size is This approach retains the multi-scale dual attention advantage of AMDT while reducing computational complexity and optimizing gradient propagation efficiency through feature splitting. The feature maps output by the CAMDT module will be used as one of the multi-scale feature outputs of the Backbone, passed to the Encoder for further semantic encoding and feature fusion, or directly participate in the multi-scale feature integration of the decoder. The robust multi-level feature representation provided by CAMDT lays a solid foundation for subsequent object detection, classification, and localization tasks, especially significantly improving the discriminative power and generalization performance of features in underwater high-noise environments.

[0052] 5) BBT module (Bayesian Enhancement Transformer)

[0053] The BBT (Bayesian Boosted Transformer) module integrates Bayesian techniques with efficient feature enhancement to construct a composite structure that balances robustness and generalization ability. Its core process first normalizes the input features (e.g., using LayerNorm or BatchNorm) to stabilize the training dynamics and accelerate convergence. Then, it extracts fine-grained spatial features step-by-step through three stacked 3×3 Dconv layers, expanding the local receptive field while reducing computational complexity and enhancing the ability to capture image details. Building on this, the module introduces a Bayesian framework, probabilistically modeling the feature transformation process through prior constraints on parameters or feature distributions (e.g., Gaussian priors). This quantifies feature uncertainty and suppresses noise interference, thereby improving the model's adaptability and generalization ability to out-of-distribution data. This module combines the efficiency of depthwise separable convolution with the statistical properties of Bayesian inference. It enhances semantic representation through multi-level spatial feature extraction and balances model confidence and generalization performance with prior knowledge. Often used as a core component of the encoder, it works in conjunction with modules like RepC3. Through cross-layer skip connections, it achieves multi-scale context fusion, further improving the network's ability to model complex scenes. See [link to relevant documentation]. Figure 4 .

[0054] The BBT-RTDETR network model includes four objective functions: localization loss GIoU (Generalized Intersection over Union) and L1 (L1 norm loss), classification loss Varifocal Loss, and Bayesian loss ELBO. The training process employs a four-stage strategy: Stage 1: Fixing the classification branch parameters, training the localization branch using only L1 and GIoU losses to optimize bounding box regression capabilities; Stage 2: Fixing the localization branch parameters, training the classification branch using only Varifocal Loss to improve classification accuracy and confidence prediction; Stage 3: Unfreezing the parameters, jointly using L1, GIoU, and Varifocal losses for model training, employing an alternating optimization strategy (odd epochs focus on localization, even epochs focus on classification) to achieve synergistic optimization of localization and classification; Stage 4: Introducing ELBO loss for Bayesian optimization based on the joint loss, while adding Gaussian noise data augmentation (σ=10,20,30) to improve the model's robustness in noisy environments and its ability to quantify uncertainty.

[0055] The localization loss function consists of L1 loss and GIoU loss, used to regress the position and shape of the target bounding box. Let the... The corresponding real bounding box is Then the loss function of L1 is defined as:

[0056] in The number of positive samples used in the loss calculation. This loss effectively reduces the distance between the predicted bounding box and the ground truth bounding box, but it is insufficient for characterizing the overall localization accuracy when the overlap is small. Therefore, the generalized IoU loss (GIoU Loss) is further introduced:

[0057] GIoU can more accurately measure the relative relationship between two boxes, especially when the predicted box does not intersect with the ground truth box, and can still provide an effective gradient signal.

[0058] Regarding the classification loss function, Varifoccal Loss is used to simultaneously optimize the class prediction probability and the target quality perception, making the class confidence learned by the model closer to the IoU-aware target quality score. Let the... The predicted class probability for each position is Its target confidence level is Then, Varifocal Loss is defined as:

[0059]

[0060] in As a balance factor, An exponential coefficient is used to adjust the difficulty of positive and negative samples. This loss not only encourages assigning higher confidence to high-quality samples (such as those with high IoU), but also effectively suppresses the impact of a large number of low-quality negative samples.

[0061] In Bayesian neural networks, the ELBO loss function from variational inference is typically used to jointly optimize model prediction performance and uncertainty representation. Its objective is to maximize the lower bound of the marginal likelihood while penalizing the difference between the predicted and prior distributions. Given data... The posterior distribution of the model parameters is a priori , If the expected value is to be determined, then the ELBO loss form used for BBT loss calculation is as follows:

[0062] The first term represents the negative log-likelihood of the data under the expected value, used to optimize the model's fit to the data; the second term is the KL divergence term, used to constrain the distribution of model parameters to not deviate from the given prior. By optimizing this loss, a predictive model with excellent performance and uncertainty quantification capabilities can be obtained simultaneously.

[0063] Step 300: Input the underwater target image to be detected into the trained BBT-RTDETR model to obtain the target detection and classification results.

[0064] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0065] experiment

[0066] The experiment compared the performance of the method of this invention with current mainstream object detection models, including YOLOv10m, YOLOv9m, YOLOv8m, YOLOv6s, YOLOv5m, Faster R-CNN, and SSD. The evaluation metric used was mean average precision (mAP).

[0067]

[0068] In summary, the experimental data fully verify the superior performance of the BBT-RTDETR method proposed in this invention for underwater noise cable detection. In particular, it significantly outperforms existing mainstream methods in terms of noise robustness, detection accuracy, and real-time performance, demonstrating high practical value and promising prospects for licensing.

[0069] Based on the same inventive concept, this application also provides a system for implementing the underwater noise submarine cable detection method described above. The solution provided by this system is similar to the solution described in the above method, and therefore will not be repeated here.

[0070] In one embodiment, an underwater noise submarine cable detection system is also provided, comprising:

[0071] The acquisition module is used to acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio.

[0072] The training module is used to input the training set into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features.

[0073] The detection module is used to input the underwater target image to be detected into the trained BBT-RTDETR model to obtain the target detection and classification results.

[0074] In the above embodiments, each module of the underwater noise submarine cable detection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0075] In one embodiment, a computer device is also provided, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps as described in all the above method embodiments.

[0076] In one embodiment, a readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps as described in all the above method embodiments.

[0077] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0078] The embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0079] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for detecting underwater noise in submarine cables, characterized in that, Includes the following steps: Acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio; The training set is input into a pre-defined BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features. The underwater target image to be detected is input into the trained BBT-RTDETR model to obtain the target detection and classification results.

2. The underwater noise detection method for submarine cables according to claim 1, characterized in that, The processing procedure of the CAMDT module includes: The input features are extracted using a series of sequentially connected AMDT modules based on a fusion channel attention mechanism and a spatial attention mechanism to obtain depth transformation features. The depth transformation features are concatenated with the input features to obtain the fused features; Channel compression is performed on the fused features to obtain the output features of the CAMDT module.

3. The underwater noise detection method for submarine cables according to claim 2, characterized in that, The AMDT module comprises, in sequence, an AMCT module, a layer normalization module, a dilated convolution module, a GeLU activation layer, another dilated convolution module, and an AMST module. The output of the AMCT module is residually concatenated with the output of the AMST module to obtain the output of the AMDT module. The processing procedure of the AMCT module includes: the original input X a After layer normalization, the input is fed into three parallel convolutional layers. The outputs of two of these convolutional layers are concatenated to obtain the first channel feature. This first channel feature is then processed by the SE attention module to obtain a channel attention weight vector. This channel attention weight vector is then multiplied channel-by-channel with the output of the third convolutional layer to obtain the feature vector. ; the features Compared with the original input X a By splicing The processing procedure of the AMST module includes: the original input X b After layer normalization, the input is fed into three parallel dilated convolutional layers, each using a different dilation rate. The outputs of two of these layers are concatenated to obtain a first spatial feature. This first spatial feature is then concatenated with the output of the third dilated convolutional layer to obtain a multi-scale spatial feature. This multi-scale spatial feature is then processed using max pooling and average pooling to obtain two spatial feature maps. These two feature maps are concatenated and processed through a convolutional layer to generate a spatial attention mask. This spatial attention mask is then multiplied pixel-by-pixel with the output of the third dilated convolutional layer to obtain the feature map. ; the features Compared with the original input X b By splicing .

4. The underwater noise detection method for submarine cables according to claim 1, characterized in that, The training process is divided into four training phases. In the first training phase, the classification branch parameters are frozen, and training is performed only by minimizing the localization loss to optimize the localization branch parameters. In the second training phase, the localization branch parameters are frozen, and training is performed only by minimizing the classification loss to optimize the classification branch parameters. In the third training phase, the parameter freeze is lifted, and the model is trained using joint loss. In the fourth training phase, Bayesian optimization is performed by introducing ELBO loss on the basis of joint loss, and Gaussian noise data augmentation is added for model training.

5. The underwater noise detection method for submarine cables according to claim 1, characterized in that, The localization loss includes GIoU loss and L1 loss, the classification loss uses Varifocal Loss, and the joint loss includes GIoU loss, L1 loss, and Varifocal Loss.

6. The underwater noise detection method for submarine cables according to claim 1, characterized in that, It also includes the following steps: Determine the evaluation metrics and evaluate the performance of the BBT-RTDETR model using a test set; The evaluation metric is the mean accuracy (mAP).

7. The underwater noise detection method for submarine cables according to claim 6, characterized in that, Gaussian noise of varying intensities was added to the images in the test set.

8. An underwater noise detection system for submarine cables, characterized in that, include: The acquisition module is used to acquire underwater target images, construct a dataset, and divide the dataset into a training set and a test set according to a preset ratio. The training module is used to input the training set into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model. The BBT-RTDETR model includes a backbone network, an encoder, a decoder, and a prediction head. The backbone network is used to extract features from the input image using multi-layer convolution and a CAMDT module to obtain multi-scale features. The encoder is used to perform layer-by-layer feature extraction and scale transformation through a RepC3 residual module, a BBT module, and an upsampling module. Simultaneously, it uses multiple skip connections to perform multi-scale semantic fusion on the multi-scale feature maps obtained by the backbone network to obtain multi-scale semantic fusion features. The decoder and prediction head are used for target detection and classification based on the multi-scale semantic fusion features. The detection module is used to input the underwater target image to be detected into the trained BBT-RTDETR model to obtain the target detection and classification results.

9. A computer device, characterized in that, Includes: memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images