Segmentation of Sidescan Sonar Imagery Using Markov Random Fields and Extreme Learning Machine

Abstract

As a widely used segmentation scheme, Markov random field (MRF) utilizes $k$-means clustering to calculate the initial model for sidescan sonar image segmentation. However, for the noise and intensity inhomogeneity nature of the sidescan sonar images, the segmentation results of $k$-means clustering have low accuracy, motivating us to use machine learning methods to initialize MRF. Meanwhile, an extreme learning machine (ELM), a supervised learning algorithm derived from the single-hidden-layer feedforward neural networks, learns faster than randomly generated hidden-layer parameters and is superior to a support vector machine (SVM). Therefore, in this paper, we proposed a novel method for sidescan sonar image segmentation based on MRF and ELM. The proposed method segments sidescan sonar images in object-highlight, object-shadow, and sea-bottom reverberation areas. Specifically, we intend to use an ELM to get an initial model for MRF. Moreover, to improve the stability of an ELM, a simple ensemble ELM (SE-ELM) based on an ensemble algorithm is utilized to obtain the prediction model. In an SE-ELM, we use an ensemble of ELMs and majority votes to determine the prediction of testing data sets. Then, the classification results of the SE-ELM are utilized to initialize MRF, termed as SE-ELM-MRF. With features consisting of pixels of small image patches, our experiments on real sonar data indicate that the SE-ELM performs better than other machine learning methods such as ELM, kernel-based extreme learning machine, SVM, and convolutional neural networks. Moreover, using SE-ELM as the initial method in the proposed SE-ELM-MRF, the segmentation results are smoother and the segmentation process converges faster than the traditional MRF.