Temporal Visual Features and Adaptive Decision Scheme for Large-scale Video Rating System

In this study, we propose three lightweight visual features and adaptive decision scheme for indecent scene detection in a large-scale video rating system. The temporal visual features consist of TMEF (Temporal Motion Energy Features), TCEF (Temporal Color Energy Features), and TCHF (Temporal Color Histogram Features). And the adaptive decision scheme uses optimal combining rule based on the properties of point estimation theory and hypothesis test. For indecent scene detection, we construct datasets of adult video segments and benign video segments with a basic unit of 30 seconds of the original movie or video track. And we use supervised learning engine of SVM (Support Vector Machine) for the performance analysis of the proposed visual features. In the experimental results, we show that TCHF and the adaptive optimal combining rule have the performance improvements in accuracy and processing time compared to the legacy MPEG-7 visual descriptors. Additionally, the proposed adaptive decision scheme can be used as a crucial piece to a robust video rating system.