Enhancing image retrieval performance by integrating BPSO with the KNN algorithm

Abstract

In this study, we propose an improved method to enhance retrieval effectiveness by combining feature selection and similarity refinement. Image features are extracted from deep learning models, including ResNet50 (RES), DenseNet201 (DEN), and DINOv2 (DINO), and are subsequently optimized using the Binary Particle Swarm Optimization (BPSO) algorithm on an evaluation set to reduce redundant features. During the retrieval stage, the proposed method employs k-nearest neighbors with label voting to identify a consistent neighborhood set, which is then used to construct a refined similarity measure. Experimental results on the Corel1K, Oxford17, and Caltech101 datasets show that the proposed method improves retrieval performance, achieving scores of 0.9950, 1.000, and 0.9166, respectively. These findings demonstrate improved retrieval effectiveness on the evaluated datasets.