Topic Analysis Based Adult Image Recognition

With the rapid development of Internet, wireless communication technology will go to the 4th generation. However, the pornographic information in the broadcast of multimedia is growing at the unprecedented speed. It is harmful to the mental health of people. How to filter the pornographic information effectively has been put on the agenda. In this paper, we systematically analyzed the advantages and disadvantages of the existing methods. And the following problems are studied: Robust skin color modeling in multi-light conditions; analysis and discrimination on pornographic images. The new theories of computer vision and pattern recognition are involved in this paper. The research is challenging and valuable in both theory and practice. The main contributions of this paper are summarized as follows:To handle the low precision of skin detection in color images, a bi-Gaussian mixture model named Bi-GMM model in YCbCr color space is proposed with two main steps. Firstly, it uses GMM model to divide the image into skin-like region and other regions. Secondly, it models the skin-like region such as hair with another GMM model and further divides the skin-like region into skin region and non-skin region.For the color cast image caused by light, a detection method based on face information is proposed, which determines the image color cast according to its color unbalance and the size of the skin region in the face region. Gray World based color correction is done in the colorcast images. Experimental results demonstrate that the proposed method can improve the accuracy of detection of skin regions effectively.The discrimination based on low-level features tends to result in mistakes. Based on image represented by bag of words model, image topic analysis model such as latent dirichlet allocation model is introduced to judge the adult images according to its latent semantic information. Experimental results prove the effectiveness of the proposed method.