A Hybrid Approach to Detect Adult Web Images

摘要

This paper presents a hybrid approach to discriminate benign images from adult images. Different from previously published works, our approach combines face detection and adaptive skin detection. First, face detection using haar-like features is performed, then skin color model is learned on-line by incorporating the information of color distribution in the face regions. Based on the result of face detection and adaptive skin detection, a set of semantically high-level features are extracted. These features have human-oriented meanings which can effectively discriminate adult images from benign images. Our two main contributions are i) a set of high-level features for adult or benign image classification and ii) a novel adaptive skin detection algorithm in still images. Experimental results are reported to demonstrate the strength of our approach.