Font Size: a A A

Research On Color Attribute And Deformable Model Fused Pornographic Image Detection Method

Posted on:2015-02-16Degree:MasterType:Thesis
Country:ChinaCandidate:D Y LiFull Text:PDF
GTID:2268330431465326Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
The fourth generation of mobile communication technology makes the Internetpermeates people’s life further. However, the spread of pornographic image in theInternet arises a lot of social problems. Therefore, the study of the detection andfiltration of the pornographic image in the network is urgent. In this paper, we train thesensitive organs model to identify the pornographic image on the basis of analyzing theexisting methods. The main research achievements are summarized as follows:Traditional pornographic image detection algorithm detects the skin region andthen analyzes it, which highly depends on the skin detection accuracy. Therefore, wepropose a pornographic image detection method based on the models of sensitive organtrained by the deformation model. The multi-scale and deformable object detectionmodel is trained based on the histogram of oriented gradients of the objects by the latentsupport vector machine. The experimental results show that the proposed methodreduces the false detection rate greatly with comparative detection rate compared withthe bag of words model on pornographic image detection.Considering the color distribution of sensitive organs in pornographic image isconsistent, we introduce the color attribute into the deformation model. Using the colorand gradient distribution characteristics to train the models of sensitive organs. Colorname is adopted to train the color attribute, which is fused with the feature based on thehistogram of oriented gradients to train the multi-scale and deformable sensitive organsmodel by the latent support vector machine. In the experiments, we test three sensitiveorgan models separately, then combine them together to judge the pornographic image.The experimental results show that compared with the bag of words model, theproposed method improves the detection rate by6%and reduces false detection rate by18%.This research combines skin color and gradient distribution characteristics, whicheffectively avoids the un-detection of sensitive organs of traditional skin color detectionmethod. Thus the detection accuracy of our method is improved and the detection erroris reduced. This paper provides a theoretical basis and efficient way for the research andapplication of pornographic image detection.
Keywords/Search Tags:HOG feature, Color attribute, Deformable model, Latent SVM
PDF Full Text Request
Related items