Abstract
Hashing methods help retrieve swiftly in large-scale dataset, which is important for real-world image retrieval. New data is produced continually in the real world which may cause concept drift and inaccurate retrieval results. To address this issue, hashing methods in non-stationary environments are proposed. However, most hashing methods in non-stationary data environments are supervised. In practice, it is hard to get exact labels of data especially in non-stationary data environments. Therefore, we propose the unsupervised multi-hashing (UMH) method for unsupervised image retrieval in non-stationary environments. Thus, in the UMH, a set of hash functions is trained and added to the kept list of hash functions sets when a new data chunk occurs. Then, multiple sets of hash functions are kept with different weights to guarantee that similarity information in old and new data are both adapted. Experiments on two real-world image datasets show that the UMH yields better retrieval performance in non-stationary environments than other comparative methods.