Classifying of Objectionable Contents with Various Audio Signal Features

Abstract

The rapid development of multimedia related technologies and internet infrastructure have made general users can create, edit, and post their contents and can easily access any content that they desire. But it also leads to the harmful side effects that are creation and uncontrolled distribution of objectionable contents. Especially it is very serious for pornographic contents that are more than about 70% of objectionable contents. The objectionable contents mean the pornographic contents in this paper. Most of the related studies are focused on image-based approaches and there are few studies based on audio-based approaches. In this paper, we try to classify objectionable contents based on various audio signal features. The audio signal features used in this paper are perceptual features that are spectral properties, MFCC based feature set and TDMFCC based feature set. For the reasonable results, we define the audio-based objectionable contents model and then construct dataset according to the defined model. For training and classifying dataset of two classes, objectionable and nonobjectionable class, SVM classifier is used. TDMFCC based feature set has a good performance of accurate rate with SVM classifier, about 95%, and the results show that it is very effective to detect and classify the objectionable contents based on audio features.