Visual search tasks to discover target in distractors are used for performance test of early vision. Fast and spatially parallel detection is taken as evidence that features in question are coded early in the visual process. Targets that are defined by conjunctions of features are usually found through a serial process of checking and rejecting distractors. The search time increases linearly with the number of distractors, suggesting that attention must be focused to each item in turn in order to conjoin features. Based on these recent studies of visual search, a model of visual attention is proposed. It is assumed that a visual image is encoded in a multi-resolution pyramid and attention function selects a sampling area from the pyramid. The function, guided by top-down and bottom-up mechanisms, give priority to sampling. New sampling is reconstructed by combining with previous samplings so that the reconstruction is as visually recognizable as possible at any moment. A computer simulation of the model produces the same general characteristics as human.