LEARNING REGULARIZED MULTI-VIEW STRUCTURED SPARSE REPRESENTATION FOR IMAGE ANNOTATION

Automatic image annotation is an important problem in computer vision owing to its critical role in image retrieval. In order to exploit the diversities of different features in a sample as well as the similarities, we present a regularized multi-view structured sparse representation model for image annotation. In this model, handcrafted visual features, deep learning based features and label information are considered as different views. Each view is coded on its associated dictionary to allow flexibility of coding coefficients from different views, while the disagreement between each view and a soft-consensus regularization term is minimized to keep the similarity among multiple views. The weight for each view is learned in the coding stage, and a weighted label prediction and propagation method is also proposed. Experimental results on ESP Game and IAPR TC-12 datasets demonstrate the effectiveness of the proposed approach compared with other related approaches for image-annotation task.