Semantic image retrieval using deep feature representation, LDA, and extreme gradient boosting

Content-based image retrieval (CBIR) remains challenging because of the semantic gap between low-level visual representations and high-level image concepts, redundancy in high-dimensional deep features, and limited discrimination among visually similar categories. This study proposes a classification-guided CBIR technique that integrates fine-tuned VGG-16 feature representation, supervised linear discriminant analysis (LDA), eXtreme Gradient Boosting (XGBoost), and cosine-similarity ranking. A fine-tuned VGG-16 network extracts 4096-dimensional semantic representations from the FC2 layer, which are projected by LDA into a compact $$C - 1$$ -dimensional discriminative space learned exclusively from the training partition. XGBoost subsequently models nonlinear category boundaries and predicts the semantic category of each query, thereby restricting the candidate retrieval space. Final ranking is performed independently using cosine similarity between the LDA-transformed query and candidate feature vectors, and the Top-20 images are returned. The proposed technique was evaluated under a stratified 80:20 closed-set protocol on Corel-1 K, Corel-5 K, Corel-1500, and OTScene. The proposed technique achieved mAP@20 values of 96.40%, 85.00%, 93.72%, and 83.96%, respectively, while maintaining an average retrieval time of 0.34 s. Controlled ablation experiments demonstrated that the proposed technique consistently outperformed eight alternative configurations, with an exact binomial sign test confirming statistical significance ( $$p=0.0078$$ , two-sided). These results show that supervised discriminant compression, nonlinear category guidance, and normalized similarity ranking provide complementary benefits for accurate and computationally efficient retrieval in labeled image collections. The proposed framework therefore offers a reproducible and effective solution for classification-guided CBIR.