Visual saliency detection based on topographic independent component analysis

A computational model of visual saliency detection is proposed based on topographic independent component analysis. This model consists of three steps: first training basis functions and extracting features which represents complex cell responses by topographic independent component analysis, then estimating feature distributions, and finally calculating self-information and obtaining saliency maps. It is demonstrated by numerical examples that the proposed model could detect saliency regions in some circumstances when previous related model couldn´t, and predict human attention fixations better than other models.