Detection of interest points in turbid underwater images

Our research is motivated by an evident lack of evaluation of recent image matching techniques for applications in underwater vision. This paper is a first step in this direction. This work compares the performance of popular salient keypoint detectors on images degraded by turbidity. We show that, as opposed to over-land, on images acquired in water medium, Hessian-based approaches outperform their Laplacian and Harris counterparts. Fast Hessian, the detector of the Speeded Up Robust Features (SURF) matching technique, is recognized to be the best method for scale-invariant detection. Conversely, when invariance to scale is not required, a combination of standard Hessian and Harris with sub-pixel accuracy and non-maxima suppression is more accurate. The objective of our work was also to create and distribute a reference set of turbid images, which can be used to evaluate processing, detection, description and matching techniques for underwater applications. We present a collection of 36 images acquired by a specially designed trinocular system under 12 gradually increasing turbidity levels. We also draw attention to image quality assessment method called SSIM, Structural SIMilarity index, which reliably quantifyes degradation of image quality caused by turbidity. As a whole, the major goal of this paper is to provide an updated reference for researchers dealing with keypoint detection in underwater imaging.