Research On Underwater Pipeline Object Detection In Side-Scan Sonar Images Based On Deep Learning

Underwater pipelines are an important means of energy transmission and information exchange,and regular inspections are required to ensure the integrity of underwater pipelines.In order to achieve online intelligent detection of underwater pipelines,this thesis proposes an object detection algorithm for underwater pipeline side-scan sonar images with high detection accuracy,high speed,and can be deployed to embeded devices.Due to the detection difficulties such as false detection and missed detection of obscure objects caused by noise similar to the object features in the image,this thesis conducts an in-depth study on the detection method of underwater pipeline objects based on deep learning of side-scan sonar images,and the main research work is as follows:(1)Establish underwater pipeline Side-scan sonar image dataset.Since there is no public dataset of underwater pipeline side-scan sonar images,this thesis obtains image data through field tests,pre-processes the acoustic images using data noise reduction and expansion techniques according to the image characteristics,and then establishes the underwater pipeline dataset by manual annotation.(2)Proposed deep learning based underwater pipeline recognition algorithm.YOLOV5 is used as the base model of the underwater pipeline recognition algorithm,and Coordinate Attention(CA)mechanism is added to the YOLOV5 feature extraction network to ensure more effective representation of location features and reduce the false detection caused by the similarity between the object and some background features.In order to make the recognition model better use features of different scales and different resolutions,the Bidirectional Feature Pyramid Network(Bi FPN)structure is used instead of the Path Aggregation Network(PANet)structure in YOLOV5,which can better fuse important features by weighted summation and reduce the misdetection caused by simple and small object features.The proposed PANet structure can better fuse the principal features by weighted summation,and reduce the missed detection caused by simple and small object features on the image.The experimental results show that the recognition algorithm proposed in this thesis improves the average accuracy by 9.2% compared with YOLOV5,which verifies the recognition performance of the underwater pipeline recognition model.(3)Proposed underwater pipeline position fitting strategy.Due to the complex underwater environment noise brought about by the false recognition,it is not possible to fit all the prediction results,so the pipeline fitting strategy with object space continuity and position constraint is established and studied in comparison with Random Sample Consensus(RANSAC)algorithm and least squares method.The experimental results show that the average error of the slope of the pipeline alignment fitted by this method is only0.00775,which effectively solves the problem of long-range anomalous data affecting the fitted pipeline alignment information.(4)Deploy the underwater pipeline object detection model integrated on the Jetson Nano board.The median filtering noise reduction operation,underwater pipeline identification model and position fitting model are ported and deployed on Jetson Nano,and the overall underwater pipeline object detection model has an average time of 0.277 s for single image detection,which verifies the feasibility of deploying the underwater pipeline object detection algorithm to Jetson Nano for object detection.