Browsing by Subject "Relevance feedback"

Finding information such as images that match the users need is a demanding task. The user has to be aware of the right keywords to search with and the underlying collection might not contain the intended images for all users. We propose a methodology to overcome these limitations by generating images matching the users needs instead of retrieving them from a collection. The methodology combines generative adversarial networks (GANs) capable of generating photorealistic images with relevance feedback from users. The methodology was evaluated with 29 user experiments where users were given a set of image retrieval tasks to find images not present in the collection of images. The images generated using the system were compared to the baseline images the users picked from a set of images representing image search results. The results from the study show that image quality was improved significantly across all image domains (cats, cars, beds, faces) and relative improvement was greater the less the starting image had in common with the target. The results suggest that generating images can be more useful than retrieving them from an existing collection.