Joint MLP and Pseudo-label Unsupervised Hash for Image Retrieval

The increasing growth of image data has made the implementation of image retrieval in big data environment a hot research topic. Traditional hashing technique uses hand-designed features for learning, which can't obtain high-level semantic features. Deep hashing has become a powerful tool for solving multimedia data retrieval in recent years with the powerful feature learning capability and the inherent advantage of small hash code storage/fast retrieval. Most of these algorithms rely on manual tagging, which is very expensive. In this paper, we propose the Joint MLP and Pseudo-label unsupervised hash (JMPH). In this approach, object features in the image are extracted with the target detection method as pseudo-labels to train the deep hash. In particular, we additionally devise the MLP network to conduct the encoding process. Experimental results on two datasets demonstrate that our unsupervised hashing method outperforms the state-of-art unsupervised hashing methods.