Two Phases Neural Network-Based System for Pornographic Image Classification
Usama Sayed, Samy Sadek, Bernd Michaelis · 2009
A robust model for skin detection is the primary need of many fields of computer vision, including f ace detection, gesture recognition, and pornography det ection. In 1996, the first paper on pornographic im age detection was published. Since then, different rese archers argue different color spaces to be the best choice for skin detection in pornography detection. Unfortunat ely, no comprehensive work attempts to use more than one color space and evaluate its performance for detect ing pornography. In this paper, a new two phases ne ural-based system for classifying images into two classes por nographic and non-pornographic is proposed. The proposed system makes use of a fast and precise neural netwo rk model based on adaptive transfer function, calle d Multilevel Sigmoidal Neural Network (MUSNN). Furthermore, the system exploits 5 color spaces in all th eir possible representations for skin detection in porn ographic images. Receiver Operating Characteristics (ROC) curve of the proposed system shows that our system outperforms other pornography detection schemes in the context of detection rate and false positive rate. Plenty of experimental results are presented includ ing photographs and a ROC curve calculated over a test set of photographs, which show stimulating performance.