Abstract: A new method for multi-label image annotation was proposed based on the combination of visual features and tag consistency. Firstly, a tag consistency model TC-KSVD was established for the training images using the KSVD method. In order to further improve the annotation accuracy, multi-view visual features were incorporated into the model. This method was arranged not only to utilize the discriminant model of the training samples ’item labels and coding coefficients, but also to utilize the relationship between tags and the coding coefficients, so as to increase the discriminability of the dictionary and improve the annotation performance. The experimental results on the Corel5K datasets show that, the VTC-KSVD method with multi-view visual features and tag consistency can accurately find the neighbors with similar visual features and semantic features, which can significantly improve the annotation accuracy and can effectively alleviate the sparsity problem caused by limited training data.