Computational Models for Top-down Visual Attention

This chapter mainly discusses the computational models by combining bottom-up and top-down processing. In the combined models, the bottom-up part in almost all models uses all or part of the core of the BS model, and the top-down part often adopts other methods in computer vision, such as neural networks.

Seven types of top-down computation are presented in this chapter. The most comprehensive is the population based model in which feature representation is in cell population form, and it is biologically plausible. Many modules are considered in the model such as bottom-up feature extraction, top-down knowledge leaning and storage, feature update by top-down influence, object recognition, inhibition of return (IoR), eye movement map, and so on. This model is first introduced in Section 5.1, then Section 5.2 covers the hierarchical object search model which simulates the human search method from coarse to finer resolution according to top-down intention. The decision tree as top-down knowledge learning, storage and retrieval is introduced in the Sections 5.3 and 5.4.

Sections 5.5 and 5.6 illustrate the two models of the simple VOCUS and the model with fuzzy ART. Finally, we introduce the top-down SUN model with Bayesian framework in Section 5.7.

The seven typical top-down computation methods combining the bottom-up model give different methods of knowledge representation, storage, learning and influence on visual attention.