Length adaptive hashing for semi-supervised semantic image retrieval

Image Hashing methods have proven to be both effective and efficient for large-scale image retrieval problem. The advances in hashing methods concentrate on the learning of image features and hash tables. Existing hashing methods manually select fixed hash code length for all classes of images in a large database. However, we have observed that the length of the hash code is essential to the retrieval performance but it is rarely studied. Short hash codes cannot preserve similarity among images well while long hash codes may lead to high storage costs. Linear search for the optimal length of hash code length is time-consuming. In this paper, a semi-supervised length adaptive hashing method (LAH) is proposed to adaptively optimize hash code lengths for different semantic image classes using a multiobjective evolutionary algorithm based on decomposition. Two objectives regarding retrieval precision and storage cost are set for optimization. We conduct experiments on three real-world image databases and the experimental results show that the proposed LAH significantly improves the retrieval performance compared to the original traditional semi-supervised hashing methods.