Automatic Detection and Classification of Man-made Targets in Side Scan Sonar Images

The automatic detection of man-made targets in side scan sonar images is of great interest in a variety of applications ranging from the detection of illegal waste disposal by passing ships to the detection of mine-like objects in military applications. The standard solution is to perform image segmentation (highlight and shadow pixels are distinguished from background pixels) followed by classification based on feature extraction. One of the problems faced during image segmentation is the need for an adaptive threshold due to the varying illumination often found in side scan sonar images. Dark and bright bands are a frequent occurrence in these images. This paper introduces a novel way to improve the contrast of a side scan sonar image and at the same time, balance the illumination throughout the image, thus eliminating the need for adaptive thresholding. The self-adaptive power filtering technique will be discussed and the interesting results from this technique will be presented. The simplicity of this technique also makes it a suitable candidate for real-time processing. Besides the usual features (e.g. size) extracted for classification, we will also be introducing contour-specific features to differentiate between objects with regular outlines (i.e. man-made objects) and those with irregular outlines (e.g. rocks). Our classification uses a divide-and-conquer approach. Highlight and shadow regions are first evaluated separately. This will allow simple features to be used more effectively. The remaining regions are then merged and new features are derived from the merged regions to further reduce the false alarm rate. Sand ripples are a source of high false alarm rate. Also, they may distort the shape of targets which may lead to a high false rejection rate. Thus ripple detection was useful. In this paper, we highlight an interesting phenomenon observed in the 2D Fourier transforms of side scan sonar images with ripples. This observation enabled us to successfully detect ripples with a high success rate.