Asymmetric hashing with multi-bit quantization for image retrieval

In the hashing approaches with multi-bit quantization, each projected dimension is divided into multiple regions indexed with multiple bits to preserve the neighborhood structure of the data. However, the query is processed in binary, and the distances between the adjacent regions are usually assumed to be equal, resulting in the accuracy loss of the computed distance between the data and the query. To tackle the above problems, in this paper an approach is proposed to process the query and the data asymmetrically. By representing the data with the expectation value of the region where the data belong and preserving the query in original form, the distance between the query and the data can be computed accurately. A specific asymmetric approach with non-parametric multi-bit quantization is further developed for the PCA (Principle Component Analysis) hashing method. With the special consideration of PCA characteristic, every projected dimension is adaptively divided into a certain number of the regions according to the minimal variance. The results of the experiments have shown that the better performance can be obtained in the asymmetric hashing approach with multi-bit quantization than that in the other approaches, and can be improved further in the specific asymmetric approach with non-parametric multi-bit quantization with respect to the PCA hashing method.