Learning Perceptual Hash Similarity for Image Copy Detection

Image copy detection is commonly addressed using either local descriptors or deep learning models, which can be computationally expensive and rely on high-dimensional features. In contrast, this work explores copy detection using compact perceptual hash representations and learned similarity functions defined directly on hash codes. We evaluate classical hash distances under realistic transformations using the PIHD dataset and assess generalization on a modified MS COCO dataset (mCOCO). We propose SiPHaD, a Siamese-based model that learns similarity in hash space, improving retrieval performance while maintaining efficiency. Results demonstrate that lightweight hash-based approaches, when combined with learned similarity, provide a strong alternative to feature-heavy pipelines.