An Interactive System based on Yes-No Questions for Affective Image Retrieval

Abstract

We propose an interactive system based on yes-no questions for affective image retrieval. We propose two querying methods, a question generation method, Affective Question and Answer (AQA), and a feedback method, Affective Feedback (AF). Conventional image search systems ask users to input queries by text. However, it is not always easy for users to convert their intention into verbal representations. Especially, the query generation becomes even more difficult when a user tries to find images with affective words due to its subjectivity. In addition, it is not guaranteed that the images are properly annotated with enough number and high quality of tags. To solve these problems, we propose a yes-no questions-based image retrieval system that can effectively narrow down the candidate images. We also provide an affective feedback interface in which users can do the fine tuning of weights of the affective words. We conducted experiments on image retrieval task with 117,866 images. The results showed that our system brings satisfactory results to users in case where the proper text querying is difficult.