Pornographic image screening by integrating recognition module and image black-list/ white-list subsystem

摘要

A number of pornographic image detection techniques have been studied in the literature but the online error-correction method is rarely discussed. As some of the pornographic images on the Internet are duplicated, the online error-correction capability can prevent repeated misclassification occurring on the same image. An image black-list/white-list subsystem is presented to correct the classification errors of the recognition module. This subsystem uses the wavelet transform to extract the colour and texture features from images. However, images can be lossily compressed and the image features will be shifted. To be able to identify lossily compressed images, a support vector machine serves as a classifier to determine whether two image feature vectors represent the same image. The experimental result demonstrates that the proposed subsystem can remedy the errors of the recognition module and increase the recognition accuracy. 1 Introduction Because of the large amount of pornographic data (text, images, videos and so on) existing on World Wide Web sites, how to block them from children?s access has become an active research topic. Most of the relative works have focused on detecting pornographic images. Lee et al. [1] used the AdaBoost method to classify images as naked or not according to the features extracted from the skin regions. A similar approach is also adopted in the work of Zheng et al. [2]. Jones and Rehg [3] detected the skin pixels by calculating the skin likelihood ratio of each pixel. Then, seven features are extracted and input into a neural network to classify the image. Shih et al. [4] and Wang et al. [5] retrieved similar images from a pre-organised image database for an unknown image. If there is a certain portion of adult images in the retrieved images, the unknown image is recognised as an adult image. Forsyth and Fleck [6] connected the edge points of the skin areas to form limbs and segments and used a set of geometric constraints to deter-nmine whether a human figure is present. The method of Yang et al. [7] extracted the regions of interest (ROI) from the image first.