Projection-based enhanced residual quantization for approximate nearest neighbor search

To reduce the time costs on approximating vector quantization of image features and training codebooks for high-dimensional vectors, a projection-based enhanced residual vector quantization was proposed. Based on previous research on enhance residual quantization (ERVQ), the principle component analysis (PCA) was combined with ERVQ, then both training codebooks and quantizing feature vectors were done in low-dimensional vector space to improve the efficiency. The features for training codebooks were projected into low-dimensional vector spaces. The overall errors generated by projection and quantization were both considered in training procedure to increase the codebook discrimination. For the proposed quantization method, a method to fast computing the approximate Euclidian distance between vectors was designed to retrieve approximate nearest neighbor exhaustively. The experimental results show that the proposed approach only takes near 1/3 training time compared to ERVQ on the condition of same retrieval accuracy. Meanwhile, the proposed approach outperforms the other state-of-the-arts over time efficiency on training codebooks, retrieval accuracy and efficiency. This study provides a reference for the effective combination of PCA with other quantization models.