With the rapid developments of machine vision and artificial intelligence, the visual attention mechanism, as an important part of machine vision, has attracted more and more attention. A coarse-to-fine saliency detection method is proposed based on non-subsampled contourlet transform (NSCT), which, as a frequency-domain based saliency detection method, can make full use of the low-frequency and high-frequency information of images and suppress the influence of illumination on detection as well. First, the non-subsampled contourlet transform is used to decompose the input images. The low-frequency components are enhanced by Retinex to ameliorate the brightness uniformity of images, and thus the influence of illumination on the saliency detection is suppressed. Then, the coarse saliency detection is performed. The high-frequency components are enhanced nonlinearly to suppress noises and enhance details, and thus the high-frequency feature maps are obtained via reconstruction. The global and local saliency analyses of the high-frequency feature maps are performed within the scope of low-frequency coarse saliency maps. Finally, the fine saliency maps are obtained via fusion. The contrast experiments are carried out on three datasets and the results confirm the feasibility and effectiveness of the proposed method.