Adult movie classification system based on multimodal approach with visual and auditory features

This paper proposes the adult movie classification system that is based on multimodal features; visual and auditory features. Visual features consist of image-based feature and video-based feature. Some MPEG-7 image descriptors and the rule of thirds are adopted as imaged-based feature for deciding harmfulness of a single video frame. The temporal color histogram feature and the repeated curve-like spectrum feature are used as video-based and audio-based feature respectively. They can detect efficiently the motion and sound properties that are appeared in most indecent scenes. In classifying a movie file, we use multi-level decision and classification; feature-level decision, clip-level decision and finally file-level classification. Support vector machine classifier is used at the feature-level decision. Each single feature-based movie classification performance has about or a slightly higher than 90% of accuracy and multimodal feature-based movie classification performance is improved up to 96.5% of accuracy under our dataset configured with 500 general movies and 500 adult movies. The measured performance shows that multimodal approach can be deployed to improve the classification performance.