Automatic Image Annotation for Semantic Image Retrieval

This paper addresses the challenge of automatic annotation of images for semantic image retrieval. In this research, we aim to identify visual features that are suitable for semantic annotation tasks. We propose an image classification system that combines MPEG-7 visual descriptors and support vector machines. The system is applied to annotate cityscape and landscape images. For this task, our analysis shows that the colour structure and edge histogram descriptors perform best, compared to a wide range of MPEG-7 visual descriptors. On a dataset of 7200 landscape and cityscape images representing real-life varied quality and resolution, the MPEG-7 colour structure descriptor and edge histogram descriptor achieve a classification rate of 82.8% and 84.6%, respectively. By combining these two features, we are able to achieve a classification rate of 89.7%. Our results demonstrate that combining salient features can significantly improve classification of images. Disciplines Physical Sciences and Mathematics Publication Details This article was originally published as Shao, W, Naghdy, G and Phung, SL, Automatic Image Annotation for Semantic Image Retrieval, Lecture Notes in Computer Science, 4781, 2007, 369-378. Copyright SpringerVerlag. This journal article is available at Research Online: http://ro.uow.edu.au/infopapers/739 Automatic image annotation for semantic image retrieval Wenbin Shao, Golshah Naghdy, and Son Lam Phung SECTE, University of Wollongong, Wollongong NSW, 2522 Australia {ws909,golshah,phung}@uow.edu.au Abstract. This paper addresses the challenge of automatic annotation of images for semantic image retrieval. In this research, we aim to identify visual features that are suitable for semantic annotation tasks. We propose an image classification system that combines MPEG-7 visual descriptors and support vector machines. The system is applied to annotate cityscape and landscape images. For this task, our analysis shows that the colour structure and edge histogram descriptors perform best, compared to a wide range of MPEG-7 visual descriptors. On a dataset of 7200 landscape and cityscape images representing real-life varied quality and resolution, the MPEG-7 colour structure descriptor and edge histogram descriptor achieve a classification rate of 82.8% and 84.6%, respectively. By combining these two features, we are able to achieve a classification rate of 89.7%. Our results demonstrate that combining salient features can significantly improve classification of images. This paper addresses the challenge of automatic annotation of images for semantic image retrieval. In this research, we aim to identify visual features that are suitable for semantic annotation tasks. We propose an image classification system that combines MPEG-7 visual descriptors and support vector machines. The system is applied to annotate cityscape and landscape images. For this task, our analysis shows that the colour structure and edge histogram descriptors perform best, compared to a wide range of MPEG-7 visual descriptors. On a dataset of 7200 landscape and cityscape images representing real-life varied quality and resolution, the MPEG-7 colour structure descriptor and edge histogram descriptor achieve a classification rate of 82.8% and 84.6%, respectively. By combining these two features, we are able to achieve a classification rate of 89.7%. Our results demonstrate that combining salient features can significantly improve classification of images.