Saliency segmentation and foreground extraction of underwater image based on localization

Saliency detection aims at automatically estimating visually salient object regions in an image, saliency segmentation and foreground extraction are two important applications of this. However, it is a challenge for underwater images to estimate salient regions by conventional saliency detection methods because of the low-contrast and poor quality. In this paper, we address this problem by combining the detected object regions rather than the whole image, where Fish Localization is used for proposing candidate regions. We extensively evaluated our method on 780 underwater images, and experimental results show that the performances of saliency detection and segmentation are improved. These saliency segmentation masks are further used to extract the foreground objects of an image. It is well proved that our approach is fast and efficient for underwater images which are low-contrast, poor quality and with multiple salient objects.