Detecting cartoons: a case study in automatic video-genre classification

Abstract

This paper presents a new approach for classifying individual video frames as being a 'cartoon' or a 'photographic image'. The task arose from experiments performed at the TREC-2002 video retrieval benchmark: 'cartoons' are returned unexpectedly at high ranks even if the query gave only 'photographic' image examples. Distinguishing between the two genres has proved difficult because of their large intra-class variation. In addition to image metrics used in prior cartoon-classification work, we introduce novel metrics like ones based on the pattern spectrum of parabolic size distributions derived from parabolic granulometries and the complexity of the image signal approximated by its compression ratio. We evaluate the effectiveness of the proposed feature set for classification (using support vector machines) on a large set of keyframes from the TREC-2002 video track collection and a set of Web images. The paper reports the identification error rates against the number of images used as training set. The system is compared with one that classifies Web images as photographs or graphics and its superior performance is evident.