Visual-attention-based thumbnail using two-stage GrabCut

This paper proposes a new thumbnail generation method to improve the recognizability of visual attention objects on small displays. Previous methods such as simple scaling reduce the recognizability of original images because the visual attention objects become too small to recognize. When we view thumbnails on small displays such as those of mobile devices, recognizability is indispensable for handling many images simultaneously. To solve the problem of low recognizability of visual attention objects, we adopt GrabCut to extract visual attention objects from an original image and then divide the original image into visual attention objects and a background image. While the background image is reduced to fit the size of a thumbnail, the extracted visual attention objects are merged into the reduced background image to preserve their recognizability. In adopting GrabCut, we propose a two-stage GrabCut method to automate the extraction of attention objects; the extraction was performed by hand in previous methods. Our experimental results show that our proposed method is able to shorten the search time by 44% and improve the precision of the search by 19% in comparison with simple scaling.