Cross-level Diffusion-based Affinity Learning for Unsupervised Salient Object Detection

Numerous bottom-up salient object detection (SOD) methods rely on local similarity (affinity) to construct, which ignores the relationships among non-adjacent regions. To solve this problem, we propose a novel framework based on absorbing Markov chain (AMC) with cross-level diffusion-based affinity learning, namely CLD. In the AMC model, high transition probability implies reliable relationships among regions. For this purpose, a dense transition probability matrix is learned by a cross-level diffusion strategy. First, to comprehensively reveal the relationship among regions, multi-level features are extracted and are used to calculate multiple initial sparse affinity matrices. Second, the tensor product in the iterative diffusion process has the property of capturing the high-order relationships among features. Thus, by iteratively updating sparse affinity matrices, multiple-level features are effectively fused. Finally, through a weight learning paradigm, outlier data is removed and noise is fully suppressed. Meanwhile, learning weights are also assigned for each sparse affinity matrix and they are integrated into a dense transition probability matrix. Furthermore, to fit salient objects at multiple scales, we propose a refinement model MCLD which is a Bayesian framework under the guidance of high-level features. The MCLD aggregates saliency maps generated by CLD at different scales, and calibrates their shortcomings to produce more stable results. Extensive experiments on five challenging datasets indicate that the proposed method outperforms ten state-of-the-art algorithms.

I. Introduction

The presence of the visual attention mechanism enables humans to quickly identify interested objects with complex external information. Salient object detection (SOD) approaches aim to realize this mechanism. As an important pre-processing method, SOD has seen widespread application in a variety of computer vision tasks, such as video processing [1], image segmentation [2], video action recognition [3], and so on. Although computer technology has made dramatic progress recently, SOD is still a challenging task.