CN-121982505-A - Submarine noise submarine cable detection method and related device

CN121982505ACN 121982505 ACN121982505 ACN 121982505ACN-121982505-A

Abstract

The invention discloses an underwater noise submarine cable detection method and a related device. The method comprises the steps of obtaining an underwater target image, constructing a data set, inputting a training set into a BBT-RTDETR model for training to obtain a trained BBT-RTDETR model, wherein a backbone network in the BBT-RTDETR model is used for carrying out feature extraction on the input image by adopting a multi-layer convolution module and a CAMDT module to obtain multi-scale features, an encoder is used for carrying out layer-by-layer feature extraction and scale conversion through a RepC residual module, the BBT module and an up-down sampling module, carrying out multi-scale semantic fusion on the multi-scale feature map by matching with a plurality of jump connections to obtain multi-scale semantic fusion features, and a decoder and a prediction head are used for carrying out target detection and classification based on the multi-scale semantic fusion features. The invention can improve the precision of submarine cable detection in the underwater noise environment.

Inventors

XIONG HUI
ZHAO ZIQIN
WEI LI
YANG CHONG
FU LIRONG
LIU JINYI

Assignees

海南大学

Dates

Publication Date: 20260505
Application Date: 20260112

Claims (10)

1. The submarine noise submarine cable detection method is characterized by comprising the following steps of: acquiring an underwater target image, constructing a data set, and dividing the data set into a training set and a testing set according to a preset proportion; The training set is input into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model, wherein the BBT-RTDETR model comprises a backbone network, an encoder, a decoder and a prediction head, wherein the backbone network is used for carrying out feature extraction on an input image by adopting a multi-layer convolution and CAMDT module to obtain multi-scale features, the encoder is used for carrying out layer-by-layer feature extraction and scale conversion by a RepC residual module, a BBT module and an up-down sampling module, and simultaneously carrying out multi-scale semantic fusion on a multi-scale feature map obtained by the backbone network by matching with a plurality of jump connections to obtain multi-scale semantic fusion features; And inputting the underwater target image to be detected into a trained BBT-RTDETR model to obtain target detection and classification results.
2. The submarine cable detection method according to claim 1, wherein the CAMDT module processing comprises: Carrying out feature extraction based on a fusion channel attention mechanism and a spatial attention mechanism on input features through a plurality of AMDT modules which are connected in sequence to obtain depth transformation features; Splicing the depth transformation features and the input features to obtain fusion features; And carrying out channel compression on the fusion characteristics to obtain CAMDT module output characteristics.
3. The submarine cable detection method according to claim 2, wherein the AMDT module comprises a AMCT module, a layer normalization module, an expansion convolution module, a GeLU activation layer module, an expansion convolution module and an AMST module which are connected in sequence, the output of the AMCT module is connected with the output of the AMST module in a residual way to obtain the output of the AMDT module, The processing process of the AMCT module comprises the steps of inputting an original input X a into three parallel convolution layers after layer normalization processing, wherein the outputs of the two convolution layers are spliced to obtain a first channel characteristic, processing the first channel characteristic through an SE attention module to obtain a channel attention weight vector, and multiplying the channel attention weight vector with the output of a third convolution layer channel by channel to obtain the characteristic And combining the features Splicing with original input X a to obtain The AMST module comprises the processing procedures that an original input X b is subjected to layer normalization processing and then is input into three parallel expansion convolution layers, the three expansion convolution layers use different expansion rates, outputs of two expansion convolution layers are spliced to obtain a first spatial feature, the first spatial feature is spliced with outputs of a third expansion convolution layer to obtain a multi-scale spatial feature, the multi-scale spatial feature is subjected to maximum pooling and average pooling processing to obtain two spatial feature images, after the two spatial feature images are spliced, a spatial attention mask is generated through the convolution layer processing, and the spatial attention mask is multiplied with the outputs of the third expansion convolution layer pixel by pixel to obtain a feature And combining the features Splicing with original input X b to obtain 。
4. The submarine cable detection method according to claim 1, wherein the training process is divided into four training stages, a first training stage, freezing of classification branch parameters, training by minimizing positioning loss only, optimizing of positioning branch parameters, a second training stage, freezing of positioning branch parameters, training by minimizing classification loss only, optimizing of classification branch parameters, a third training stage, parameter freezing removal, model training by joint loss, a fourth training stage, bayesian optimization by introducing loss ELBO on the basis of joint loss, and model training by adding gaussian noise data enhancement.
5. The submarine noise sea cable detection method according to claim 1, wherein the positioning loss comprises GIoU loss and L1 loss, the classification loss uses Varifocal Loss loss, and the joint loss comprises GIoU loss, L1 loss and Varifocal Loss loss.
6. The submarine cable detection method according to claim 1, further comprising the steps of: Determining an evaluation index, and performing performance evaluation on the BBT-RTDETR model through a test set; The evaluation index is an average precision average mAP.
7. The submarine cable detection method according to claim 6, wherein gaussian noise of different intensities is added to the test set image.
8. An underwater noise submarine cable detection system, comprising: The acquisition module is used for acquiring an underwater target image, constructing a data set and dividing the data set into a training set and a testing set according to a preset proportion; the training module is used for inputting the training set into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model, wherein the BBT-RTDETR model comprises a backbone network, an encoder, a decoder and a prediction head, the backbone network is used for extracting characteristics of an input image by adopting a multi-layer convolution module and a CAMDT module to obtain multi-scale characteristics, the encoder is used for extracting the characteristics layer by layer and converting the scales by the RepC residual module, the BBT module and an up-down sampling module, and simultaneously carrying out multi-scale semantic fusion on a multi-scale characteristic map obtained by the backbone network by matching with a plurality of jump connections to obtain multi-scale semantic fusion characteristics; The detection module is used for inputting the underwater target image to be detected into the trained BBT-RTDETR model to obtain target detection and classification results.
9. A computer device comprising a memory for storing a computer program, and a processor for implementing the method according to any one of claims 1 to 7 when the computer program is executed.
10. A readable storage medium, characterized in that it has stored thereon a computer program which, when executed by a processor, implements the method according to any of claims 1 to 7.

Description

Submarine noise submarine cable detection method and related device Technical Field The invention relates to the technical field of target detection, in particular to an underwater noise submarine cable detection method and a related device. Background The underwater target detection technology is one of core technologies in the fields of ocean engineering and resource development, and the application scene of the underwater target detection technology covers submarine cable inspection, pipeline maintenance, marine organism monitoring and the like. Due to complex optical scattering, low contrast and high noise interference (such as sediment suspension, biological disturbance and equipment noise) existing in an underwater environment, the detection accuracy of the traditional computer vision-based method (such as edge detection and threshold segmentation) is obviously reduced due to insufficient feature extraction capability. In recent years, deep learning-based object detection models (such as YOLO series (You Only Look Once), fast R-CNN (Faster regional convolutional neural network, fast Region-based Convolutional Neural Network)) have improved detection performance through end-to-end feature learning, but image degradation caused by underwater noise (such as gaussian noise, pretzel noise and motion blur) still causes a dramatic decrease in model generalization ability. Aiming at the problem of noise robustness, the existing research mainly adopts a data enhancement (such as noise injection, GAN (generation countermeasure Network), GENERATIVE ADVERSARIAL Network generation countermeasure training) or a self-adaptive method, but has the problems of large deviation between simulation noise and real environment, redundancy of a model structure and the like. The transducer architecture exhibits advantages in complex scene modeling due to its global attention mechanism, such as DETR (detection transducer, detection Transformer) series models that enable end-to-end detection through encoder-decoder structures, but it is computationally complex and sensitive to local features. RTDETR as a real-time detection model optimizes the reasoning speed, but still faces the challenges of insufficient characteristic interaction capability and missing noise suppression mechanism under underwater noise interference. In addition, the Bayesian deep learning method improves the uncertainty perception capability of the model by introducing probability modeling, and the image reconstruction technology combined with Gaussian process regression (Gaussian Process Regression, GPR) can effectively compensate the semantic information of a noise area, but the existing research has not fully fused the method into a lightweight detection framework. How to construct a detection system which gives consideration to real-time performance, noise robustness and embedded deployment requirements becomes a technical bottleneck for intelligent underwater submarine cable detection. Disclosure of Invention In order to solve the technical problems, the invention provides an underwater noise submarine cable detection method and a related device, so that submarine cable detection under underwater noise can be well performed, and submarine cable detection accuracy under underwater noise environment is improved. In order to achieve the above purpose, the technical scheme of the invention is as follows: an underwater noise submarine cable detection method comprises the following steps: acquiring an underwater target image, constructing a data set, and dividing the data set into a training set and a testing set according to a preset proportion; The training set is input into a preset BBT-RTDETR model for iterative training to obtain a trained BBT-RTDETR model, wherein the BBT-RTDETR model comprises a backbone network, an encoder, a decoder and a prediction head, wherein the backbone network is used for carrying out feature extraction on an input image by adopting a multi-layer convolution and CAMDT module to obtain multi-scale features, the encoder is used for carrying out layer-by-layer feature extraction and scale conversion by a RepC residual module, a BBT module and an up-down sampling module, and simultaneously carrying out multi-scale semantic fusion on a multi-scale feature map obtained by the backbone network by matching with a plurality of jump connections to obtain multi-scale semantic fusion features; And inputting the underwater target image to be detected into a trained BBT-RTDETR model to obtain target detection and classification results. Preferably, the processing of the CAMDT module includes: Carrying out feature extraction based on a fusion channel attention mechanism and a spatial attention mechanism on input features through a plurality of AMDT modules which are connected in sequence to obtain depth transformation features; Splicing the depth transformation features and the input features to obtain fusion features; And carrying out channe