Large scale reverse image search - A method comparison for almost identical image retrieval

In this paper, we presented our study and benchmark on Reverse Image Search (RIS) methods, with a special focus on finding almost similar images in a very large image collection. In our framework we concentrate our study on radius (threshold) based image search methods. We focused our study on perceptual hash based solutions for their scalability, but other solutions seem to give also good results. We studied the speed and the accuracy (precision/recall) of several existing image features. We also proposed a two-layer method that combines a fast but not very precise method with a slower but more accurate method to provide a scalable and precise RIS system. MOTS-CLES : recherche d’images inversé, pHash, optimisation, SI images