Research On Face Detection Of Erotic Image

With the development of Internet, netnews has become a familiar and convenient information source. But much pornographic and other undesirable information has seriously disturbed people's normal life on the net. Because images contain more information than texts, compared to texts, pornographic images are even more harmful. Thus recognition of pornographic images is the key of the system with filtering pornography online. In order to recognise the pornographic images, the first step is to determine the existence of the human body. So detecting human face is the primary task. In addition location face exactly can also provide important information for later tasks such as identification of face expression and detection of human body posture. Therefore, automated detection of face has become the pivotal step to analyse pornographic images, and it has far-reaching social significance.In order to study the problem of face detection in the pornographic images, this thesis initially discusses the the existing face detection methods, and then aiming at the complex model about human face in pornographic images, a hierarchical framework for face detection is designed. The framework has used several kinds features to detect face, such as color, Haar-Like features, edge shape, gray information. The main work of this paper is as follows:(1) skin segmentation is the pretreatment step for face detection. The parameters of skin Gaussian model is determined and the model based on the color space of YCbCr is established. In this paper, a color segmentation algorithm based on the K-means clustering method is proposed. This algorithm can change the original image into skin binary image. Through calculating the number of holes and the area, a candidate block is selected which generally contain face. The process of skin segmentation reduces the space which the later steps of face detection would search for.(2) In this paper, a face detection method based on Haar-Like features and Adaboost is implemented and applied into the pornographic images. The experimental results show that this method is feasible and effetive for positive and slight sheltery face. Considering original adaboost arithmetic costs too much time for training weak classifiers, a recurrence formula is deduced to calculate the false positive rate. This formula can accelerate the training speed obviously.(3) Aiming at undetected images, this paper build an average face template through clipping some face samples from pornographic images. And then hough transform and template matching method are used to locate the position where face would be and verify whether the probable position is face indeed. The final detection result will be obtained and the detection rate is improved.