Real-Time Detection Method for Submarine Pipeline Leakage Based on Deep Learning and Jetson nano

Abstract

There are many pipelines under the sea to transport oil and natural gas. However, some pipelines maybe leakage under long time exposing to the marine environment, which cause pollution and economic losses. In this paper a new object detection model is presented based on YOLOv5. Several models YOLOv5s, YOLOv5m, YOLOv51, YOLOv5x are trained for pipeline leakage dataset, the detection accuracy reaches 89.0%, 94.2%, 92.2%, 94.4%. Detection speed tested on computer can reach 97 FPS, 74 FPS, 51 FPS, 33 FPS, respectively. In order to satisfy the requirement of practical application, YOLOv5s is transplant to Jetson nano embedded system, and the detection speed is only 4.64 FPS. To enhance the response speed, we present a new model called YOLOv5-mv3s, which replaces backbone network of YOLOv5 with MobileNetV3s. The size of this model reduces to 7.4 MB, which is 53.4% of the lightest YOLOv5s model. The detection accuracy is about 92.9% with a faster detection speed of 9.8 FPS on Jetson nano. When the input image size is set to $416 \times 416$, the detection speed can reach about 11.5 FPS, while detection accuracy is about 91.6%. Compare with the existing detection method, the new lighter model is effective and applicable.