Image relevance model
Topics: AI (Deep Learning), Image Search, Ranking, Scoring, User Signals
The patent describes a system for training and using image relevance models to improve the accuracy and efficiency of image search results. For each unique image search query, an independent relevance model is trained using content features from images deemed relevant or irrelevant to past queries. This model then assigns relevance scores to new images, allowing for ordered, accurate search results. The training involves positive images (frequently selected for given queries) and negative images (selected for other queries), significantly enhancing the precision of relevant image searches.