There have been many works using deep learning models to predict the saliency map of an image. This project studies how to use different methods and the pros and cons of each method under different data sets. And based on MobileNet, a fast and effective method is proposed. Use MobileNet to extract features, and then multiply and link with the visual prior layer to obtain a visual saliency map. This method only needs to use 10% of the computing resources of the original method to achieve similar estimation results.