Consumer photo management and browsing facilitated by near-duplicate detection with feature filtering

Near-duplicate detection techniques are exploited to facilitate representative photo selection and region-of-interest (ROI) determination, which are important functionalities for efficient photo management and browsing. To make near-duplicate detection module resist to noisy features, three filtering approaches, i.e., point-based, region-based, and probabilistic latent semantic (pLSA), are developed to categorize feature points. For the photos taken in travels, we construct a support vector machine classifier to model matching patterns between photos and determine whether photos are near-duplicate pairs. Relationships between photos are then described as a graph, and the most central photo that best represents a photo cluster is selected according to centrality values. Because matched feature points are often located in the interior or at the contour of important objects, the region that compactly covers the matched feature points is determined as the ROI. We compare the proposed approaches with conventional ones and demonstrate their effectiveness.