Automatic image annotation based on Gaussian mixture model considering cross-modal correlations

摘要: A unified framework is developed for image annotation based on GMM and MB.Automatic transformed into a graph partitioning problem.GMM fitted by the RPEM algorithm utilized to estimate initial annotations.Integrating label visual similarities of images associated with labels.Max-bisection implemented rank-two relaxation heuristics. Automatic has been an active topic research in field computer vision pattern recognition decades. In this paper, we present new method automatic Gaussian mixture model (GMM) considering cross-modal correlations. To be specific, first employ rival penalized expectation-maximization (RPEM) posterior probabilities each keyword. Next, similarity constructed weighted linear combination seamlessly integrating information from both low level features high semantic concepts together, which can effectively avoid phenomenon that different same candidate annotations would obtain refinement results. Followed heuristics over built applied further mine correlation so as capture refining results, plays crucial role retrieval. The main contributions work summarized follows: (1) Exploiting trained images. (2) corresponding labels. (3) Refining set generated through solving max-bisection graph. Compared current competitive SGMM-RW, achieve significant improvements 4% 5% precision, 6% 9% recall Corel5k Mirflickr25k, respectively.