Pornographic Image Detection Based On Color Attention Mechanism

With the development of the information and Internet technologies, people canaccess useful information conveniently. But the spread of pornographic data seriouslydisrupt the network experience of users and endanger the teenagers’ mental and physicalhealth, which arise serious social problem. Thus, pornographic image filteringtechnology has become an important and urgent research topic, which is the focus ofmany researchers. This work focus on pornographic image filtering technology, whichinvolves the related theories in computer vision and statistical machine learning areas.This work can provide technical support to prevent the spread of adult image. The mainachievements of this paper are summarized as follows.1. Skin detection by maximally stable extremal region analysis algorithm detectsskin region on the skin distance map, in which the texture map is based on the Euclidiandistance. To solve this problem, texture map based on optimal neighbor differencemeasurement is introduced to the maximally stable color region algorithm in this paper.The proposed algorithm was tested over the MCG-Skin database, which outperformsthe maximally stable extremal region analysis based skin detection algorithm by4%ontrue positive rate with the8.2%false positive rate.2. Because of the complexity of pornographic images, low level feature basedpornographic filtering approaches suffer from high false positive rate. Bag-of-words is apromising model for pornographic filtering because of its superiority on semanticrepresentation. Since the scale invariant feature transform (SIFT) features used bybag-of-words model does not include color information, which is an important cue ofpornographic images, we propose to introduce color attention to bag-of-wordsrepresentation of pornographic images by three ways. The proposed algorithm wastested over the homebrew database. The true positive rate of our algorithm is improvedfrom80%to93.2%over the traditional bag-of-words based algorithm under falsepositive rate of20%.3.To avoid the dependency on codebook of bag of words method and endow thestronger semantic information to detection templates, we propose a highly efficienttemplate matching based approach to detect pornographic images. The algorithm runs ahigh throughput template matching process using a large number of manually labeledtemplates. The algorithm subsequently builds a bagging-based classifier to detect pornographic images. The proposed algorithm was tested over the homebrew database,which improves the true positive rate from80%to91.5%over the bag-of-words basedalgorithm under false positive rate of20%.