Large scale image understanding with non-convex multi-task learning
Liang Li, Chenggang Clarence Yan, Xing Chen, Shuqiang Jiang, Seungmin Rho, Jian Yin, Baochen Jiang, Qingming Huang · 2014
Large scale image understanding is drawing more and more attention from the researchers and industry. Inspired by the game theory and machine learning algorithm, this paper proposes a semantic dictionary to solve the key problem of visual polysemia and concept polymorphism in the large scale image understanding. The semantic dictionary characterizes the probability distribution between visual appearances and semantic concepts, and the learning of semantic dictionary is formulated into a minimization problem of the payoffs, where the players adjudge their strategies (i.e. the probability distribution) at each iteration. Non-convex multi-task learning is introduced to solve the above optimization problem. Finally, the wide applications of semantic dictionary are validated in our experiments, including the large scale semantic image search and image annotation.