Deep neural networks (DNNs) have been extensively applied in image processing, including visual saliency map pre-diction of images. A major difficulty in using a DNN for visual saliency prediction is the lack of labeled ground truth of visual saliency. A powerful DNN usually contains a large number of trainable parameters. This condition can easily lead to model over-fitting. In this study, we develop a novel method that over-comes such difficulty by embedding hierarchical knowledge of existing visual saliency models in a DNN. We achieve the objective of exploiting the knowledge contained in the existing visual sali-ency models by using saliency maps generated by local, global, and semantic models to tune and fix about 92.5% of the parame-ters in our network in a hierarchical manner. As a result, the number of trainable parameters that need to be tuned by the ground truth is considerably reduced. This reduction enables us to fully utilize the power of a large DNN and overcome the issue of over-fitting at the same time. Furthermore, we introduce a simple but very effective center prior in designing the learning cost function of the DNN by attaching high importance to the errors around the image center. We also present extensive experimental results on four commonly used public databases to demonstrate the superiority of the proposed method over classical and state-of-the-art methods on various evaluation metrics.