Abstract: Since most existing approaches of saliency detection generate incomplete and biologically implausible saliency map,in this paper,we proposed a saliency detection method based on multiscale analysis in frequency domain.Firstly,we used the wavelet transform to decompose the discrete cosine transform(DCT)coefficients of the input image into multi-scale amplitude spectra,and calculated the multi-scale saliency maps in the spatial domain.Then,we selected the better saliency maps according to the saliency evaluation function.Finally,we obtained a final saliency map of the input image by adaptive weighted integration of the selected multi-scale saliency maps.Experiments were conducted on different types of datasets,including psychophysical pattern,human eye fixation datasets,and salient target segmentation datasets(including ASD and ECSSD).The method achieves high accuracy on objective evaluation such as P-R curve,ROC curve,AUC index and calculation speed of these datasets.Experiments results show that the method outperforms other conventional approaches of saliency detection.