Efficient Top-k Content-Based Image Retrieval Exploiting Matrix Division Toward SIFT

With the growing demand for efficient content-based image retrieval (CBIR) systems, scalable and lightweight retrieval techniques have become essential. To obtain accurate retrieval results, several feature extraction techniques such as SIFT and deep feature vector exploiting CNN have been proposed. However, the computational cost for such techniques is enormous and the performance enhancement is still challenging. Thus, in this paper, we propose a novel method, Rank-based Feature on Grid-partitioning of SIFT Keypoints (RF-GSK), which enables effective top-k image retrieval without relying on deep learning models. RF-GSK utilizes handcrafted SIFT keypoints and transforms them into compact frequency and rank-based feature vectors through spatial grid partitioning which divides the spatial domain of an image into $m \times m$ cells. We also present how to determine the decompose parameter m by mathematical analysis adopting a quality assessment method for clustering and verify it empirically. To show effectiveness of our proposed method, we conduct the statistical analysis of RF-GSK and the state-of-the-art algorithms including memory requirement, feature size and the number of FLoating OPerations (FLOPs) as well as evaluate top-k accuracy of them. Our experimental results demonstrate that our proposed method show reasonable accuracy compared to the state-of-the-art algorithms but it is more efficient than the state-of-the-art algorithms in the aspects of memory usage, storage requirement and computational cost.