Vision-based attention in maritime environments

摘要

This paper presents a saliency inspired visual attention technique for maritime scenes. The main focus is on finding regions in images which there is a high likelihood of a maritime object being present. Experimentation has shown that many regional and global features are required because no single feature can reliably detect these objects. Examples of the features used are right angle corner detectors, edge density, and colour difference. A Gaussian classifier has been used to produce an Attention Map of pixel responses. Experiments using ground truthed images show the technique is effective on a large set of images of maritime scenes and is better at detecting maritime objects than existing generic salient detectors.