A Comparative Study of Deep Learning-Based CBIR Frameworks: From CNN Baselines to Explainable and Hybrid Models

Deep learning has proven to be abreakthrough technology transformingContent-Based Image Retrieval (CBIR). Inthis paper, we present a comparative studyof three novel frameworks of CBIR thatwere developed in a series of studies,namely a baseline of Modified CNN model,the Explain CBIR-Net utilizing explainableAI with Grad-CAM, and HybridCBIRNet,which incorporates CNN andTransformer-based architecture withweighted feature fusion. We quantitativelyestimate all models' accuracy, precision,recall, and explain ability over the MiniImageNet dataset. The findings of ourcomparative study underscore the balancebetween accuracy, interpretability, andfeature richness. The findings validate thesuperior efficacy of HybridCBIRNetwhile highlighting the value of contextualand explainable modelling towards criticalretrieval applications.