Research On Sparse Learning Theory Based Automatic Annotation For Large-scale Social Images

In recent years,with the rapid advance in digital photography and Internet technology,massive image resources have been shared by Internet users to the social image-sharing platforms.And the large amount of these social images brings great challenges to image management,query and annotation,which promotes image annotation and retrieval technology become one of the research hotspots in computer field.Nowadays,the tags of social images tend to be incomplete,inaccurate and fuzzy,which makes the description of image information inaccurate and affects the accuracy of image retrieval.Therefore,how to fill and correct user-provided tags has become a hot topic in the field of image retrieval.The automatic image annotation algorithms being proposed by previous researchers still have insufficient in image annotation accuracy and the scale of the problem.A deep research and discussion is made in this paper based on matrix completion theory in Sparse Learning field.This paper first solves the priori information fusion problem of the existing automatic image annotation using the traditional matrix completion model.So the inherent sparsity of the image-tag matrix is introduced,and it makes use of the prior consistency principle of the image-to-visual content and the semantic relevance between tags.The problem of automatic annotation of social images is modeled as a regularized low-rank matrix completion model combining priori sparse information.Then the alternating direction method of multipliers(ADMM)in machine learning is adopted to solve the proposed model.The simulation results show that the proposed algorithm preferably solves the incomplete and noisy tags of social images,thus improving the accuracy of image retrieval.In addition,it is difficult to apply to the large-scale problem because of the fact that alternating direction method of multipliers is still inherently serial optimization algorithm,so this paper introduces parallel multi-block ADMM and stochastic gradient descent method to improve the solving efficiency of this model,and further an automatic annotation algorithm based on parallel multi-block ADMM for large-scale social images is proposed,which preferably solves the large-scale problem of automatic image annotation.