Multi-scale visual attention and saliency modelling with decision theory

Recently, an information-based saliency technique which is biologically plausible and computationally feasible called Discriminant Saliency (DIS) has been proposed. While DIS successfully defines discriminant saliency in the information theoretic sense, its implementation restraints the sampled features to a single fixed-size window and creates a bias towards objects with distinctive features fitted in the window size. This paper proposes a multi-scale discriminant saliency (MDIS) technique for visual attention which uses the wavelet transform for the multi-resolution framework. MDIS utilizes mutual information between classes and feature distribution to quantify classifying discriminant power as saliency value in multiple dyadic-scale structures. The paper will present simulations on Neil Bruce's image database with quantitative and qualitative results showing the advantages of MDIS over DIS. For quantitative comparisons, numerical tests AUC, NSS, LCC are measured and several plots are generated to visualized differences between simulation modes; meanwhile, qualitative evaluation is a visual examination of synthesized saliency maps of general natural scenes.