High-Speed Similarity-Based Image Retrieval with Data-Alignment Optimization Using Self-Organization Algorithm

In a retrieval system for vast amounts of image data, the primary storage cannot hold all image feature vectors because a huge data capacity would be required. Therefore, it is necessary to handle a slow-access secondary storage effectively as well as a first-access primary storage. In this paper, we propose a data-alignment optimization method in the secondary storage to access fast. Our idea is based on the premise that similar data have a high probability of simultaneous access in a similarity-based retrieval system. Then, we approximate the secondary storage by one dimensional alignment and allocates similar data to closer regions in the alignment. It is regarded as an optimization problem for a one-dimensional alignment that enables a fast retrieval by reducing a seek range. Through the experiments for 5 millions dataset, we confirmed that the data access time was reduced by 25.3% compared to no data-alignment optimization, which contributed significantly to speeding up the entire retrieval process.