Most of current content-based image retrieval methods are based on the stationary data environment, which is impractical. How to retrieval relevant images in non-stationary data environments has attracted more and more attentions in recent years. A few methods have been proposed to solve this problem. Incremental hashing (ICH) is a recently published method, which has achieved high retrieval performance in non-stationary data environments with concept drifts. In ICH, multiple hash tables are trained incrementally and independently based on different chunks of data arriving at different times, one after another. The similarity between some images may be not preserved well by existing hash tables. These images are regarded as badly hashed but ignored in the subsequent training of new hash tables in ICH. The “badly” hashed images may be outliers, but may also be “inliers” of new classes hence ignoring them risk losing valuable information. In this paper, the issue of how to deal with such “badly” hashed images is studied. A new method, Complementary Incremental Hashing with query-adaptive Re-ranking (CIHR), is proposed, where a new hash table is trained to be complementary to previous hash tables based on these “badly” hashed images and the newest chunk of data. Moreover, a query-adaptive re-ranking method is also proposed, which evaluates each hash function according to the position of the query with respect to the corresponding hash hyperplane. Experimental results on 15 simulated non-stationary data scenarios show that the proposed CIHR method achieves higher retrieval accuracy than comparable methods, thus setting a new state of the art in image retrieval in non-stationary data environments.
|Journal||IEEE Transactions on Multimedia|
|Publication status||Accepted/In press - 8 May 2020|