3

Computational Models in the Spatial Domain

The phenomenon of visual attention in human vision and its background of biology on anatomy, physiology and psychology are presented in the two previous chapters of Part I. In Chapter 2 we introduced several conceptual models to simulate the phenomenon of visual attention: the model of feature integrated theory (FIT), the guide search model (GS), the synchronous oscillation model and so on. Although these models are partly concordant with biological results, there is still a large gap towards practical applications. In Part II, some computational visual attention models, that have explicit mathematical formulae and are programmable in computers, are introduced. Part II includes four chapters (Chapters 3-6), covering bottom-up attention models in spatial domain and in frequency domain in Chapters 3 and 4, respectively, and attention models with top-down cues in Chapter 5. Finally, in Chapter 6 we present several criteria for measuring these computational models. In all the computational models, the result of computation appears on a topographical image (like the activation map in GS2) called the saliency map, which was first proposed by Koch and Ullman in 1985 [1]. The location of a peak on the saliency map indicates a focus of visual attention. In this chapter, several typical computational models [2–10] are introduced.