(Національний університет "Києво-Могилянська академія", 2024) Kubytskyi, Volodymyr; Panchenko, Taras
The rapid growth of image data globally has amplified the demand for effective image similarity detection methods, particularly in tasks like image deduplication. This paper introduces a novel approach using enriched image embeddings derived from combining outputs of intermediate layers of pre-trained CNNs. The proposed method improves F1 scores across tasks such as near-duplicate detection, multi-angle view analysis, and schematical layout comparisons. Real-world applications in the real estate domain demonstrated fewer errors and enhanced performance, offering a promising direction for addressing complex image comparison challenges.