Improvement and Evaluation of Visual Saliency based on Information Theory

Abstract

Visual saliency is a well-known image processing technique based on theory about visual attention of human beings. Several proposed computing approaches based on findings in psychological and neurological researches about the part of human brain area in charge of visual attention. Contrast to those methods, in this paper the information theory is recommended as backbone for a new approach of building visual attention computer model, assumed that the uniqueness of a pixel or a group of pixels correlates with saliency. The information saliency is identified in both spatial domain and temporal domain, and eventually those two pieces of information is combined in a mathematical integrated framework naturally. Experiments of the model on intensity, contrast, and especially full motion video show that its performance is comparable to other state-of-art saliency approaches. Though there are still minor flaws in the information based saliency, it is a potential alternative approach for visual attention in video-based applications.