Sequential conditional entropy maximization semi-supervised hashing for semantic image retrieval

Hashing has been widely applied to large scale semantic image retrieval. Unsupervised hashing cannot work well for semantic image retrieval while supervised hashing requiring full label information for large databases is not practical. Semi-supervised hashing (SSH) solves this problem by learning the semantic information from a small portion of labeled images and the data structure information from the unlabeled images. The major drawback of the current SSH is that they cannot guarantee the maximization of entropy over all hash bits for a better coding efficiency. We propose a SSH which maximizes the conditional entropy of a bit with respect to all previous bits, i.e. the sequential conditional entropy maximization SSH. It is further extended to a nonlinear SSH with a new mapping method to enhance precision and recall rates. Experimental results show that the nonlinear SSH does not work well for all cases, and a heuristic guideline for the selection between linear and nonlinear hashing is given based on the characteristics of the database. We also propose a multiple hashing version of the proposed method for high recall rate with few hash bucket visits. Experimental results show that the proposed method outperforms current SSH methods.