Information Gain Product Quantization for Image Retrieval

Approximate nearest neighbor (ANN) research has become the key method of content-based image retrieval. Product quantization (PQ) is a popular method for approximate nearest neighbor search. However, the informativeness may not be uniformly distributed across the subspace in product quantization. Allocating the same number of bits for the subspace will bring large quantization loss. To address this issue, in this paper, we propose an improved product quantization method, which adaptively allocates different numbers of bits to subspace via information gain. The key of our method is to find the optimal bit allocation strategy. To this end, our method maximizes the summation of information gain for each subspace, and also takes the limited length of codes into account. Experimental results on two large-scale bench-marks GIST1M and 22K-LabelMe demonstrate that our approach significantly outperforms state-of-the-art methods in terms of accuracy.