Image auto-annotation is a basic and challenge task in the image retrieval work. The traditional machine learning methods have btained a lot achievements in this field. The deep learning algorithm has achieved great success in image and text learning work since it is presented, so it can be an efficient method to solve the semantic gap problems. Image auto-annotation can be decomposed into two steps, that is, the basic image auto-annotation based on the relationship between image and tag, and the annotation enhanced based on the mutual information of the tags. In this article, the basic image auto-annotation is viewed as a multi-labelled problem. Therefore the prior knowledge of the tags can be used as the supervise information of the deep neural network. After obtained the image tag s, the dependent relationship of the tags is used to improve the annotation result. Finally, the model is tested in Corel and ESP datasets, and results prove that the method can efficiently solve the image auto-annotation problems.