Study On The Segmentation And The Detection Algorithms Of Side-scan Sonar Image

With people's increasing demand for underwater activities,side scan sonar systems on underwater vehicles have gradually become an important tool for underwater exploration and are widely used in terrain mapping,resource detection,underwater search and rescue,object detection,etc.However,there are several problems in real-time data transmission and manual object detection.First,with low power and low bandwidth,it is time consuming and energy consuming for the underwater system to transmit data.Second,since the acoustic image is annoyed by deformation and noise,the manual detection calls for rich experience.Therefore,the automatic object detection for side scan sonar data has become an important technology.Large amount of data,severe deformation and strong noise makes it very difficult for the underwater platform to achieve accurate real-time detection in the side scan process.In this paper,we divide the detection system into three parts: coarse segmentation,object detection,and accurate segmentation.We segment different areas in the sonar image fast,and then implement object detection to screen suspected objects,and finally obtain the target boundary with good noise resistance through accurate segmentation.This makes the real-time object detection achieve the satisfied accuracy.This article mainly includes the following innovations:Firstly,our method focuses on the problem that the sonar image has strong noise.Based on the graph theory method,this paper innovatively proposes a rapid bothway spanning forest algorithm for raw sonar data,which can reduce the similarity between the objects and the surrounding noise,thus the algorithm can obtain an accurate segmentation.The algorithm uses the maximum/minimum spanning tree algorithm to segment the image at the same time.After obtaining the over-segmentation result,the linear structure detection and historical information growth are used to obtain the segmentation with good integrity of large-scale object structure and effective separation of small-scale objects.Secondly,to lower the false alarm rate caused by noise,two main characteristics are proposed to detect the objects of rough segmentation: object saliency and object size.First,we use a multi-resolution statistical difference test algorithm to measure the saliency of the object.Compared to the single-scale test,the multi-resolution target detection can not only restore the target's saliency of visual effects,but also take into account the original sonar,so as to reduce the misjudgment of the noise.Secondly,we detect by estimating the reasonable target area size.Thirdly,to address the problem of low accuracy of sonar image segmentation,this paper proposes an updated level set initialization method based on the bothway spanning forest algorithm.It also allows the algorithm implemented on the local data.With better initialization,the level set curves rapidly evolve to the required boundaries,reducing the iteration number,thereby speeding up the running time and obtaining the precise edge position of the object.In the end,we build up the user interaction system and verify our algorithm on this platform with the database.The experiment shows satisfying real-time performance(1 second/ping),and get a good recall rate(93%).The false alarm rate is below 50%(48%),and the final segmentation also reach good accuracy.