High-Dimensional Image Retrieval via Adaptive Subspace Dimension Product Quantization

Abstract

To address low search accuracy and long search time in product quantization (PQ) algorithms for large-scale datasets, the Adaptive Subspace Dimension Product Quantization (ASDPQ) algorithm is proposed. It optimizes subspace partitioning by adaptively choosing the number of sub-spaces based on quantization error comparison, speeding up the search. During training, it uses two encoding patterns and selects the optimal one for efficient quantization. A high-dimensional data image retrieval model is developed. In experiments on SIFT and GIST ANN search datasets, ASDPQ outperforms OPQ and PQ algorithms, with recall rates of 0.84 and 0.97, and search times of 3.135ms and 5.374ms respectively. It also reduces addition computation by 4.54% and 6.96% compared to PQ. When integrated into an image retrieval system, it achieves a similarity rate of over 80% and an average shortest retrieval time of 2.63ms, demonstrating its effectiveness and reliability in high - dimensional data image retrieval.