Published Paper Details:

CBIR (Content-Based Image Retrieval) is significant in various applications, including digital asset management, medical diagnosis, and surveillance, in which images must be retrieved based on visual content. Conventional CBIR methods rely on handcrafted features or CNN when essentially low-level spatial information is acquired, while the model may not describe the long-range dependencies and contextual semantics. Recent progress with Transformer-based models has shown enhanced contextual representation; however, these models typically ignore pixel-level features, which may provide incomplete representations and degrade retrieval performance. To overcome those limitations, this study presents HybridCBIRNet, a new hybrid deep learning framework that leverages the advantages of CNNs and Transformers via a weighted feature fusion mechanism. The CNN part is meant for hierarchically extracting spatial features, and the Transformer module provides global contextual information. The complementary representations are fused to produce a rich and discriminative feature embedding. Moreover, the system incorporates an Explainable AI (XAI) module that offers image retrieval interpretability, making it more applicable in sensitive areas. The proposed framework is validated on the Mini ImageNet dataset. It shows the best performance compared to the other existing methods with an accuracy of 97.25%, precision of 96.80%, recall of 97.10%, and mAP of 97.00%. These findings confirm the strength of the hybrid feature representation and verify the model's capacity to harvest highly similar images. That is because accuracy and robustness across different categories of images are essential features for them. HybridCBIRNet improves long-tail classification, which can act as a solution to real-world applications of CBIR that require accuracy and transparency.