In the current saliency detections based on deep learning, how to make full use of the convolution features at all levels is the key issue. In order to solve this problem, we propose a saliency detection method based on full convolution neural network, which is a fusion of all convolutional features. Firstly, all the convolution features are mapped to multiple internal scales, and the saliency maps are predicted by combining the convolutional features of each level on each scale. Then the fused saliency maps are obtained by fusing the saliency maps of each scale. Finally, smooth saliency maps and optimized salient boundaries are obtained through full connected conditional random fields. Experimental results show that the proposed method has higher accuracy, recall rate and lower average absolute error in ECSSD database and SED2 database, and provides more reliable pretreatment results for target recognition, machine vision and other applications.