부산대학교 도서관

As an icon is a means of communication, understanding its concept is of utmost importance. However, in many cases, an icon’s appearance is not in line with its concept. To predict the abstract concept with appearance, we propose compositional learning for icons. The whole icon’s concept is predicted by combining its compositional components. We propose Jobicon, ajob symbolizing icon dataset suitable for compositional learning. Jobicon is classified by the components of the job’s icons. When applied in the image retrieval model, our dataset demonstrated the best performance and is suitable for compositional learning.