Combining Deep Feature and Multi-label Classification for Semantic Image Annotation

To bridge the semantic gap between data of difference modalities,a multi-label semantic image annotation approach combining deep CNN(convolutional neural network)and ECC(ensemble of classifier chains)is proposed.The approach is composed of two stages,i.e.generative feature learning and discriminative semantic learning.Deep CNN is utilized to extract high level visual features of images in the first step.Then ECC is trained based on the obtained visual features and image semantic label set.Meanwhile,semantic information in visual features is learned.Eventually,automatic annotation on unseen images can be done by the trained model.Experimental results on Corel5K and PASCAL VOC 2012 image datasets show that the proposed approach is more robust and more accurate than some state-of-the-art approaches.