The design of the effective feature vectors is the key to the saliency detection algorithm, which determines the upper bound of the model effect. A new saliency detection algorithm based on global model and local search is proposed by combining the deep convolution neural networks and the hand-crafted features. In the global model, the initial saliency map is generated from designing the extra convolution layers for VGG-16 network training, and thus the saliency value of each object candidate region can be predicted from a global perspective. In local optimization model, the super-pixel region with multi-degree segmentation is described by designing the contrast descriptors and region characteristic descriptors, and the saliency score of each region is predicted. Finally, a linear fitting method is used to fuse the result generated from two models, and the final saliency map is obtained. Contrast experiments for four data sets are demonstrated and the results show that the proposed algorithm has the highest precision.