Research On Submarine Cable Target Detection Technology Based On Improved YOLOV3 Algorithm

With the rapid growth of the world’s population,the human demand for resources is growing,and land resources are being depleted at a high rate,so there is an increasing demand for the development of marine resources.The submarine cable is an essential link in the development of marine resources and the island economy,but the submarine environment is very complex,and events such as sea bed scouring and ship anchor damage can cause irreversible damage to the cable.In order to ensure the normal operation of the submarine cable,it is particularly important to conduct regular inspections.With the continuous development of underwater technology and the increasing demand for inspection efficiency,the inspection of submarine cables has changed from the traditional frogman dive inspection to the intelligent underwater robot inspection.Machine vision plays a vital role in underwater robotic inspection.At present,underwater robot inspection of submarine cable through machine vision has become a technological trend,but there are still the following problems: The motion of the underwater robot causes the captured image of the submarine cable to be blurred,which in turn makes the image of the submarine cable lack information on the location and characteristics of the cable;the underwater light is dim and the water as a medium for light propagation makes the image blue-green.To address these problems and the characteristics of the submarine cable ontology as a linear object and a large percentage of the image,this paper proposes an improved target detection model YOLO-CABLE based on the YOLO-V3 algorithm,with the following main work and technical innovations:A simulated submarine cable experimental platform is built based on a real submarine environment,an underwater vehicle with an underwater HD camera is used to acquire submarine cable image data,and a diverse submarine cable dataset is created to provide data for submarine cable target identification studies.The submarine cable image dataset in this paper contains 3104 images with different disturbances,such as motion blur,partial absence,occlusion,absorption,and scattering effect of water on light.Before feature extraction,we propose to add an image pre-processing module based on data enhancement,which effectively improves the problem of difficult feature extraction due to the blue-green color of underwater images of submarine cable;in the backbone feature extraction network and feature fusion network,we propose to add skip connection and multi-structure multi-size feature fusion,which solves the problem of insufficient location information and feature information of the target of the submarine cable due to the blurred image;according to the characteristics of the submarine cable in slightly larger proportion of the image,the proposed lightweight prediction network shortens the detection time of the model and can meet the standard of real-time detection underwater;a grayscale extraction-based submarine cable angle recognition algorithm is proposed,and the submarine cable angle information is the necessary data for the underwater robot to perform the submarine cable tracking task.Based on the above research work,a target detection model for submarine cables is proposed and named YOLO-CABLE to fill the gap of the current research technology for submarine cable detection at home and abroad.The model adds skip connection in the backbone feature extraction network to enhance the location information of submarine cable;adds a top-down downsampling structure in the multi-scale special fusion to reduce the network computation and expand the network perception field;and lightens the processing in the prediction network to speed up the network detection.In this paper,the effectiveness of these improvements is demonstrated by ablation studies.When compared with other algorithms,the average detection accuracy of the YOLOCABLE model is improved by up to 4.2% and the average detection speed is reduced by up to 1.616 seconds.In the experiments of submarine cable angle identification,the accuracy of submarine cable angle identification reaches 83%.The experiments demonstrate that the YOLO-CABLE model proposed in this paper has outperformed other existing network models in the field of machine vision for submarine cable identification.