KIDS-NUTN at ImageCLEF 2012 Photo Annotation and Retrieval Task
Been-Chian Chien, Guanbin Chen, Li-Ji Gaou, Chia-Wei Ku, Rong-Sing Huang, Siao-En Wang · 2012
Abstract. The task of visual concept detection, annotation, and retrieval using Flickr photos at ImageCLEF 2012 was organized as two subtasks: concept an-notation and concept retrieval. In this paper, we present the effort of KIDS lab for the two subtasks. The proposed approaches combine various visual and tex-tual features, dimension reduction methods, the random forest classification models, and the semi-supervised learning strategy. For the concept annotation subtask, the annotation results show that combination of tags and visual features outperforms visual-only features while using the same classification model. The results also show that semi-supervised learning is not superior to supervised learning in this subtask. Further, it does not seem able to gain more advantage on F-measure when more different visual features were used. For the concept retrieval task, the results illustrate that the textual features contain much richer informatics than visual features in general retrieved concepts.