PQH: A Multithreaded Parallel NN Search Index for Content-based Image Retrieval

In this paper, we propose a NN search index method called Parallel Quantized Hashing (PQH) for relevant retrieval in Content-Based Image Retrieval (CBIR) systems. Aiming at improving the performance of query, the PQH partitions the original dataset into several subsets and keeps the pointers to quantized vectors of each subset in a hashing like structure. It only needs to filter parts of the query space to get candidates via neighboring masks and accesses the corresponding real vectors for precise calculation. The neighboring masks are sets of mask vectors indicating cells close to the query point using different metrics. The PQH is multithreaded designed so that the construction and the query can be run in parallel upon multi-core processors. Our extensive experiments confirm that the sequential PQH is of high query efficiency and outperforms the existing methods such as VA+-file and LSH. The parallel run PQH achieves even higher performance and returns appropriate NN results. And more important it is simple for practical implement in real applications.