Objectionable audio content recognition based on in-Class Clustering method

This paper focuses on automatic adult video sequences recognition from the perspective of feature porno-sounds detection, which serves as a verification step, a supplementary method and an independent detector. To the special of erotic sounds, their feature analysis is given. Our statistics and experiments show that features such as energies in subbands, µ -spectral centroid, mean of short-time zero-crossing rates, and High short-time zero-crossing rates ratio play important roles in discriminating erotic audio files. At the same time due to the complexity of data within and outside erotic audio class, in-Class Clustering is proposed which selects the most representative subclass for training and classification. All these efforts increase the recall rate and decrease the false positive rate. Experiments on real data from the Internet indicate that the proposed method yields superior performance that 85.35% recall rate and 15.46% false positive rate are achieved.