Computer vision community has long attempted to automatically detect locations in the image that are able to capture attentions of users. In recent years, more and more researchers have proposed to address this problem from the perspective of simulating human visual attention mechanisms. In this paper, we study modeling visual attention in frequency domain. Our major contributions are twofold: 1. A new method called band-divided method (BDM) is developed to generate the saliency map by integrating the amplitude spectrum with the phase spectrum. 2. A quantitative measurement according to min-distance dissimilarity (MDD) is presented to evaluate the saliency map, which is more appropriate for non-binary ground-truth data. Experiments on benchmark dataset and comparisons with traditional approaches demonstrate the promise of the proposed work.