A robust front page detection algorithm for large periodical collections

Large-scale digitization projects aimed at periodicals often have as input streams of completely unlabeled document images. In such situations, the results produced by the automatic segmentation of the document stream into issues heavily influence the overall output quality of a document image analysis system. As a solution to the issue segmentation problem, this paper introduces a robust, two-step front page detection algorithm. First, the salient connected components from the front page of the periodical are described using a multi-dimensional Gaussian distribution based on discrete cosine transform (DCT) features. Second, a graph model is computed by applying Delaunay triangulation on the selected set of components. A specialized, error-tolerant graph matching algorithm is used to compute the distance score between the model and each candidate page. Experiments on a large, real-world newspaper data set demonstrate the generality and effectiveness of the proposed method.