I. Introduction
Saliency detection is a promising research area, which is regarded as a preferential allocation of computational resources [1]–[5]. It has been applied to a wide range of visual tasks, such as image retrieval [6], image compression [7], enhancement [8]–[10], foreground annotation [11], image retargeting [12], [13], and pedestrian detection [14]. In recent years, co-saliency detection has become an emerging issue in saliency detection, which detects the common salient regions among multiple images [15]–[18]. Different from the traditional single saliency detection model, co-saliency detection model aims at discovering the common salient objects from an image group containing two or more relevant images, while the categories, intrinsic characteristics, and locations of the salient objects are entirely unknown [19]. The co-salient objects simultaneously exhibit two properties, i.e. 1) The co-salient regions should be salient with respect to the background in each image, and 2) All these co-salient regions should be similar in appearance among multiple images. Due to its superior expansibility, co-saliency detection has been widely used in many computer vision tasks, such as foreground co-segmentation [20], [21], object co-localization and detection [22]–[24], and image matching [25].