A Simulated User Study of Image Browsing Using High-Level Classification Teerapong Leelanupab, Yue Feng, Vassilios Stathopoulos, Joemon M. Jose University of Glasgow, Glasgow, G12 8RZ, United Kingdom, {kimm,yuefeng,stathv,jj}@dcs.gla.ac.uk Abstract. In this paper, we present a study of adaptive image browsing, based on high-level classification. The underlying hypothesis is that the performance of a browsing model can be improved by integrating high- level semantic concepts. We introduce a multi-label classification model designed to alleviate a binary classification problem in image classifica- tion. The effectiveness of this approach is evaluated by using a simulated user evaluation methodology. The results show that the classification assists users to narrow down the search domain and to retrieve more relevant results with respect to less amount of browsing effort. 1 Introduction The accumulation of large volumes of multimedia data, such as images and videos, has led researchers to investigate indexing and search methods for such media in order to render them accessible for future use. Early, Content Based Image Retrieval (CBIR) systems were solely based on low-level features extracted from images inspired by developments in image processing and computer vision [3]. Nevertheless, due to the “semantic gap” [13] problem, using just low-level descriptors will not lead to an effective image retrieval solution. Recent research in multimedia indexing has investigated automatic annotation methods to index multimedia data with keywords which convey the semantic content of media. Those keywords cannot however represent all aspects of image content due to inherent complexity of multimedia data. In both of the above cases in multimedia retrieval, the search paradigm is similar and inspired by traditional information retrieval systems. The searcher poses a query to the system, which can be a rough sketch: a predicate query such as “images with at least 80% blue” or a textual query, and then the system returns a ranked list of potentially relevant images. An alternative search paradigm that suits better to the nature of multimedia data, and especially images, is browsing. A well studied browsing approach is to visualize retrieved images as a graph where nodes are images and paths are relationships between them based on some underlying similarity. Browsing is facilitated by allowing users to browse the collection by following paths in this graph (e.g. [5,6]). In this approach, relevance feedback and the Ostensive Model of developing information needs can be easily integrated [15]. Browsing models are inherently different than traditional image retrieval sys- tems since the focus is not on user queries but in the user’s browsing path where