In order to solve the problems of incomplete feature learning and unobvious salient regions in complex scenes for the traditional saliency detection methods，a saliency detection algorithm based on multi-level deep fea-tures and random walk was proposed． Firstly，the fully convolutional networks (FCN) was used to perform multi-level convolution deep feature extraction on the image by combining the deep and shallow feature information． Se-condly，superpixels segmentation on the image was carried out and the extracted deep convolution features were assigned to the corresponding superpixels to construct a feature matrix． Finally，the final saliency map was generated by regularizing the random walk ranking model． The experimental results on the ECSSD and DUT-OMRON databases show that the proposed method has certain advantages in accuracy and F value，compared with the six representa-tive saliency detection algorithms．