Interpretable Region-Attribute Graph Matching Framework for Structure-Aware and Explainable Content-Based Image Retrieval

Abstract

Abstract In recent years, Content-Based Image Retrieval (CBIR) has advanced significantly with Deep Learning (DL), yet most of the existing models rely on global descriptors and attention-based local features, which struggle to capture explicit structural relationships within images. These models often operate as black boxes and provide limited interpretability; they exhibit reduced robustness under viewpoint and geometric variations. To address these limitations, an Interpretable Region-Attribute Graph Network (IR-AGN) is proposed in this research. Specifically, each image is decomposed into semantically meaningful regions, where each node encodes visual and semantic descriptors while edges represent geometric, appearance, and contextual relations. Consequently, a joint node-edge graph matching strategy computes similarity by aligning visual descriptors and relational structures between query and retrieval images. The results of the proposed IR-AGN model exhibit a higher mean Average Precision (mAP) of 95.19% on the Revisited Paris (RParis) dataset, when compared with the existing Multi-Layer Orientation Histogram (MLOH) model.