Multi-hash ok-means for large scale image retrieval problems

Abstract

Hashing methods have been widely applied to deal with large scale image retrieval problems. The orthogonal k-means (OKM) generates hash codes for images via a minimization of quantization errors using a coordinate descent algorithm. However, its performance is dependent on the initialization of the rotation matrix. Therefore, the Multi-Hash OK-means (MHOK) method is proposed to reduce the instability of performance and improve the precision and recall rate yielded by the OKM. The MHOK uses multiple hash tables created by multiple of OKM with different initialization. For each image, instead of keeping all multiple OKM hash codes, the MHOK selects a subset of hash codes via a minimization of quantization errors over different hash tables. Therefore, hash codes of different images are independently selected from different hash tables. Thus, the overall quantization error of all images is minimized over the multiple OKM hash tables. So, the MHOK enjoys the high performance of multi-hashing while uses less storage by hash code selection. Experimental results show that the MHOK achieves better precision-recall curve than the original OKM and other state-of-art unsupervised hashing methods.