Image Saliency Estimation Via Random Walk Guided by Informativeness and Latent Signal Correlations
Jia Li,Shu Fang,Yonghong Tian,Tiejun Huang,Xiaowu Chen
DOI: https://doi.org/10.1016/j.image.2015.05.005
IF: 3.453
2015-01-01
Signal Processing Image Communication
Abstract:Visual saliency is an effective tool for perceptual image processing. In the past decades, many saliency models have been proposed by primarily considering visual cues such as local contrast and global rarity. However, such explicit cues derived only from input stimuli are often insufficient to separate targets from distractors, leading to noisy saliency maps. In fact, the latent cues, especially the latent signal correlations that link visually distinct stimuli (e.g., various parts of a salient target), may also play an important role in saliency estimation. In this paper, we propose a graph-based approach for image saliency estimation by incorporating both explicit visual cues and latent signal correlations. In our approach, the latent correlations between various image patches are first derived according to the statistical prior obtained from 10million reference images. After that, the informativeness of image patches and their latent correlations are jointly considered to construct a directed graph, on which a random walking process is performed to generate saliency maps that pop-out only the most salient locations. Experimental results show that our approach achieves impressive performances on three public image benchmarks.