Abstract

In this paper, we describe an approach to camera metadata estimation using regression based on Independent Component Analysis (ICA). Semantic scene classification of images using camera metadata related to capture conditions has had some success in the past. However, different makes and models of camera capture different types of metadata and this severely hampers the application of this kind of approach in real systems that consist of photos captured by many different users. We propose to address this issue by using regression to predict the missing metadata from observed data, thereby providing more complete (and hence more useful) metadata for the entire image corpus. The proposed approach uses an ICA based approach to regression.