Vector Quantization for Large Scale CBIR

In order to meet the requirements of image retrieval and real-time index storage of massive datasets, a Content-based Image Retrieval (CBIR) based on vector quantization is used in this paper. By the feature extraction of depth residual network, the improved Product Quantization (PQ) coding algorithm is used to generate index and dynamic database under FAISS Framework, which can realize nonlinear large-scale CBIR with high robustness. Empirical results show that the CBIR based on improved PQ has achieved good performance in index compression rate, recall rate and retrieval time compared with PQ and Locality-Sensitive Hashing (LSH) in SIFT1M dataset.