Active Learning for Detection of Mine-Like Objects in Side-Scan Sonar Imagery

Abstract

Abstract-A data-adaptive algorithm is presented for the selection of the basis functions and training data used in classifier design with application to sensing mine-like targets with a side-scan sonar. Automatic detection of mine-like targets using side-scan sonar imagery is complicated by the variability of the target, clutter, and background signatures. Specifically, the strong dependence of the data on environmental conditions vitiates the assumption that one may perform a priori algorithm training using separate side-scan sonar data collected previously. In this paper, a novel active-learning algorithm is developed based on kernel classifiers with the goal of enhancing detection/classification of mines without requiring an a priori training set. It is assumed that divers and/or unmanned underwater vehicles (UUVs) may be used to determine the binary labels (target/clutter) of a small number of signatures from a given side-scan collection. These sets of signatures and associated labels are then used to train a kernel-based algorithm with which the remaining side-scan signatures are classified. Information-theoretic concepts are used to adaptively construct the form of the kernel classifier and to determine which signatures and associated labels would be most informative in the context of algorithm training. Using measured side-looking sonar data, the authors demonstrate that the number of signatures for which labels are required (via diver/UUV) is often small relative to the total number of potential targets in a given image. This procedure designs the detection/classification algorithm on the observed data itself without requiring a priori training data and also allows adaptation as environmental conditions change.

Image Processing Techniques For the Detection and Classification of Man Made Objects in Side-Scan Sonar Images

“…The advantages for relevance vector machines over support vector machines is the availability of probabilistic predictions, using arbitrary kernel functions and not requiring to set many parameters. Details of the implementations and results can be found in Dura et al (2005)Couillard et al (2008 It is important to highlight, that is not always necessary to use all the features; sometimes using a smaller is better than using a large set of features which are correlated. Therefore some all the supervised techniques reviewed used some optimisation procedures before the training process to determine the best combination of features.…”

“…In the context of side-scan sonar this is very important as labeling the data is very expensive, a diver or unmanned underwater vehicle with a camera has to label it. An active-learning algorithm based on semi-supervised techniques was first proposed by Dura et al (2005). The algorithm, kernel-based, was developed with the goal of enhancing mine detection/classification of mines without requiring a priori data set.…”

A Side Scan Sonar Image Target Detection Algorithm Based on a Neutrosophic Set and Diffusion Maps

“…Reed et al [6] introduced an automatic mine-like object detection method for SSS images, and the detection phase was completed with an unsupervised Markov random field (MRF) model, but the MRF is parameter-based, and the physical size and geometric signature of the mines were also required for accurate mine-like target detection. Dura et al [7] proposed a target detection method based on sparse expression; however, this method requires an unmanned underwater vehicle (UUV) to provide a small number of samples with indicator classes, such as target and background, in the current measurement area. Grasso et al [8] proposed a small target detection method based on local gray level information and a mathematical morphology operation, but the method only fitted those SSS images obtained by an automated underwater vehicle (AUV) platform.…”

Active Learning: Any Value for Classification of Remotely Sensed Data?

“…Thus, it potentially leads to greater information exploitation for the data and significant reduction of the annotation cost. Recently, AL has gained attention for classification of remotely sensed data (a survey in [12]), and has also been investigated for image segmentation [13], target detection [14] and regression [15]. The focus of this paper is to provide an overview of active learning strategies for supervised classification, summarizing some of the most popular approaches, then providing more details on methods that have been developed recently to address specific challenges and opportunities in analysis of remote sensing images.…”