Gaussian mixture background for salient object detection

Salient object detection has become a valuable tool in image processing. In this paper, we propose a novel approach to get full-resolution saliency maps. The input image is segmented into superpixels, each of them presents an irregular but homogenous area of the image thus can be treated as an image unit. Intuitively, superpixels touching the image borders will have the potential to capture the background information. Therefore, pixels belong to those superpixels are collected as background samples to train a Gaussian mixture model. The saliency of each superpixel is then defined by computing the weighted probability density of the Gaussian mixture model followed by an enhancement and smoothness step. At the end, a dense conditional random field based refinement tool or cellular automata is selected by an adaptive threshold to remove the false salient regions or find other potential saliency regions to get a more accurate result in pixel-level. We compare our method to five saliency detection algorithms which are classic or similar to ours but published in recent years on a commonly used challenging dataset ECSSD. Experiments show that our approach outperforms others well.