Underwater salient object detection by combining 2D and 3D visual features

Abstract Automatic salient object detection over images is an important step for object detection and recognition. Many existing salient object detection methods are performed excellently on ground, but, it remains a challenge to detect salient objects in water. Different from common scenes on ground, challenges to underwater salient object detection are posed by poor underwater image quality, uncontrolled underwater objects and environments. Till now, most of existing methods are hindered from adaptability to difficult underwater environments, due to the strong light attenuation and scattering effects. To this end, we propose a novel underwater salient object detection method by combining 2D and 3D visual features. For feature combination, a novel detection model is established by mathematically stimulating the biological vision mechanism of aquatic animals. This model, apart from the 2D visual features i.e. the color and intensity, extracts 3D depth features to arouse the depth sensitivity in the three-dimensional space. These 2D and 3D visual features are combined by our biologically inspirited model, generating comprehensive salient object detection results. Here, aiming to correctly estimate 3D depth features from underwater images, a regional method is used to respectively extract 3D depth features in two regions, namely the artificial light and natural light regions. Evaluations show the diverse and comprehensive performances of various features for underwater salient object detection. High accuracy of our proposed method is demonstrated by comparing to state-of-the-art saliency detection methods on public underwater benchmarks acquired in diverse underwater environments.