| Summary: | 博士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 100 === In this dissertation, we propose a novel personalized ranking system for amateur photographs. Our goal of automatically ranking photographs is not intended for award-wining professional photographs but for photographs taken
by amateurs, especially when individual preference is taken into account. Photographs are described using 20 image features which can be categorized into three types: photo composition, color and intensity distribution, and features for personal preferences. We adopt RBF-ListNet as the ranking algorithm. RBF-ListNet is based on an efficient algorithm, ListNet, using radial basis functions. The performance of our system is evaluated in terms of Kendall’s tau rank correlation coefficient, precision-recall diagram,
and binary classification accuracy. The Kendall’s tau value (0.434) is higher than those obtained by ListNet and support vector regression (SVR). The precision-recall diagram and binary classification accuracy (93%) is close to the best results to date for both overall system and individual features. To realize personalization in ranking, we propose three approaches: feature-based, example-based, and list-based approach. User studies indicate that all three approaches are effective in both aesthetic and personalized ranking. In particular, the example-based approach obtained the highest user experience rating among all three.
|
|---|
