Near-duplicate image detection for the insurance field

Near-duplicate detection aims at finding images from the same scene or place inside a wide collection. In this context, near-duplicate (ND) detection differs from the extensively studied image retrieval task because of its objective is finding different version of a same object in which the distinction between ND and non-ND pairs is done through a classification function while image retrieval is usually a search to retrieve semantically similar images based on a specific query. Near-duplicate discovery has many applications as social media evaluation, web content creation, image forensic and fraud detection. I created a machine learning method that exploits images' descriptors to learn similarities between image pairs. In particular, I developed a Siamese architecture composed by convolutional neural networks to extract image pairs' features and calculate similarity scores and pairs' distances. The dataset used to train the network is a private database owned by Reale Mutua Assicurazioni, formed by claims mostly regarding building damages. The final goal is to identify all possible near-duplicates, thresholding the output score of the model, starting from a set of queries. Because false positives have to be checked by the user, a high specificity is required. To evaluate results the problem is a binary classification of near-duplicate vs. not-near-duplicate pairs. Metrics are evaluated through Receiver Operating Curve (ROC) and a final Area Under the Curve score of 0.73 is achieved.