A Robust Neural System for Objectionable Image Recognition

Abstract

A reliable model for human skin is a significant need for a wide range of computer vision applications ranging from face detection, gesture analysis, content-based image retrieval systems, searching and filtering image content on the web, and to various human computer interaction domains. In this paper, a robust neural model for human skin recognition is first presented. Then, a fully automated neural network based system for recognizing naked people in color images is proposed. The proposed system makes use of a fast and precise neural model, called Multi-level Sigmoidal Neural Network (MSNN). Furthermore, the system exploits four different color models in all their possible representations to precisely extract color features from skin regions. Receiver Operating Characteristics (ROC) curve illustrates that the proposed system outperforms other stat-of-the-art schemes of objectionable image recognition in the context of detection rate and false positive rate. Abundance of experimental results are presented including test images and the ROC curve calculated over a test set, which show stimulating performance of the proposed system.