Saliency detection via alternative optimization adaptive influence matrix model

Salient object detection based on unsupervised learning method has received widespread attention. Most researches are focused on feature selection for classifying salient objects and background. However, as saliency is difficult to describe with fixed feature, these methods can not cover all possibilities adequately. In this paper, we propose a novel method for saliency optimization by alternative optimization adaptive influence matrix (AO-AIM). First, we propose a model which directly optimizes original saliency map given by any methods, and use edge information to avoid poor feature expression. Second, AO-AIM model optimizes parameters adaptively by analyzing the relationship between regions in each image, without calculating feature's weight by statistic learning in advance. Third, we propose an alternative optimization structure, by alternative operating adaptive influence matrix and geodesic weighted Bayesian model, over-modifying causing by single-method iteration is avoided, and AO-AIM model gives a better performance. Using existing methods as prior distribution, AO-AIM model achieves significant improvement. Experiments on ASD, CSSD, SED2 and DUT-OMRON demonstrate that AO-AIM model can provide an obviously optimization on saliency detection.