Superpixel attention guided network for accurate and real-time salient object detection

Edge information has been proven to be effective for remedying the unclear boundaries of salient objects. Current salient object detection (SOD) methods usually utilize edge detection as an auxiliary task to introduce explicit edge information. However, edge detection is unable to provide the indispensable regional information for SOD, which may result in incomplete salient objects. To alleviate this risk, observing that superpixels hold the inherent property that contains both edge and regional information, we propose a superpixel attention guided network (SAGN) in this paper. Specifically, we first devise a novel supervised deep superpixel clustering (DSC) method to form the relation between superpixels and SOD. Based on the DSC, we build a superpixel attention module (SAM), which provides superpixel attention maps that can neatly separate different salient foreground and background regions, while preserving accurate boundaries of salient objects. Under the guidance of the SAM, a lightweight decoder with a simple but effective structure is able to yield high-quality salient objects with accurate and sharp boundaries. Hence, our model only contains less than 5 million parameters and achieves a real-time speed of around 40 FPS. Whilst offering a lightweight model and fast speed, our method still outperforms other 11 state-of-the-art approaches on six benchmark datasets.