In this paper we investigate the use of content-basedudimage descriptors for enhancing the performance of personudannotation in personal photo management applications. Theuddescriptors examined are related to the context of personudrecognition through face and body-patch feature matchingudin personal digital photos. In order to identify the best performing content-based descriptors, we first study a number of colour and texture descriptors for body-patch matching and face recognition descriptors for face matching using a suitably chosen data set taken from typical personaludphoto collections. We then analyse the performance of threeuddifferent fusion schemes to identify the best combination ofudcolour, texture and face recognition descriptors. Finally, we apply those descriptors to the problem of person annotation and measure their performance using a test data set, which comprises 7 different real-life personal photo collections. The experimental results illustrate that combining bodypatch feature matching with face recognition significantly improves the performance of person annotation. We further show that combining colour with texture leads to improved performance of body-patch matching. The content-based image descriptors identified in this paper show great potential for person annotation in personal photo management applications.
展开▼
