Classification and mosaicking of side scan sonar image

Abstract

As an underwater detection sensor, side-scan sonar plays an important role in marine survey, mineral exploration, underwater archaeology and so on. During the use of side-scan sonar, classifiication and mosaicking of collected images is essential in most cases. There are two main contributions in our work. On the one hand, we propose a supervised learning method based on kernel-based extreme learning machine (KELM) to perform image classification. As a single-hidden layer feedforward neural network, ELM has one hidden layer and one output layer. It has been proved that ELM provides better performance in classification and regression at shorter consumed time than some others, such as traditional support vector machine (SVM), without complex parameter adjustment. However, the weights of ELM hidden layer are randomly produced and the classification results of ELM are different because of this. To solve this problem, the kernel-based ELM was proposed, in which the hidden layer was processed with a kernel function to eliminate randomness. On the other hand, the side-scan sonar images and the classified image data will be geo-referenced mosaicked using positions produced by extended Kalman filter (EKF) with sensor data from an autonomous underwater vehicle (AUV). To eliminate gaps in the mosaicking images, image dilation is adopted in our work. Experimental results demonstrate that the proposed classification method works well, and the proposed image mosaicking method is applicable when concerns real side-scan sonar images.