<p>Near-duplicate image (NDI) detection under photometric and geometric variations remains challenging in large-scale visual retrieval. We propose a lightweight, training-free hybrid approach that integrates Scale-Invariant Feature Transform (SIFT) with DCT-based Perceptual Hashing (PH). SIFT captures distinctive, geometrically invariant local features, while PH generates compact hash codes robust to photometric distortions, together providing complementary strengths for reliable NDI detection. Our methodology introduces an algorithmic framework that computes independent similarity scores from both techniques and fuses them at the score level. Extensive experiments are conducted on the Airbnb near-duplicate dataset (Boston, Berlin, and Seattle) as well as public benchmarks including INRIA Holidays and California-ND. We report standard dataset-level classification metrics (Accuracy, Precision, Recall, F1-score, ROC-AUC) and include representative deep embedding baselines using Contrastive Language-Image Pre-training (CLIP) for comparison. Across datasets, the proposed fusion framework consistently improves robustness relative to standalone perceptual hashing while remaining competitive with SIFT-based matching. On the California-ND benchmark, the calibrated fusion model achieves 95% accuracy and an F1-score of 0.948, demonstrating strong robustness under heterogeneous photometric and geometric distortions. For comparison with modern deep embedding approaches, the CLIP baseline (pretrained on over 400 million image–text pairs) achieves higher absolute accuracy (e.g., 0.69 on Airbnb and 0.99 on California-ND), but requires heavier deep-learning inference, thus serving as a strong deep-learning reference rather than a directly comparable classical baseline.</p>