Research On The Related Technologies Of Image Filtering Based On Skin Detection

With the development and popularization of the Internet, It gave rise to unprecedented information and facilitated people's lives. At the same time, the negative information on the Internet has also brought a huge negative impact to human. In particular, it has brought serious harm to the physical and mental health of young people. In order to prevent its further diffusion, at present, many scholars engaged in research in this area. In light of the performance of negative information on the Internet, researching on network filtering technology is no longer confined to using Web site blocking and sensitive keywords matching technology, we must consider the combination of image filtering techniques to prevent the spread of unhealthy information. Founded on this, this article focuses on the study of the key technologies of Content-Based erotic image filtering made by our predecessors, and on this basis, we construct and realize an effective erotic image filter based on skin detect. The major research work is as follows:(1) We build an image database, including the test image database and the marked skin-mask image database, which contains different color, different numbers and different body posture images, and these images provide the experimental evidence for subsequent research.(2) We have studied and analyzed the three current common models of skin detection algorithm, such as the Chroma Space Algorithm, GMM Algorithm and Histogram Algorithm, and compared the three kinds of skin-color detection algorithm. In light of the shortcomings in Histogram Model, we improved it. The results show that the improved skin detection method is superior to the traditional Histogram Model in the positive detection rate and the false detection rate.(3) In the parts of feature extraction and classification, this dissertation studies and analyzes feature extraction based on the mask image. We have extracted 12 features here. Through the statistics and analysis of these features in the standard mask Image Library, this dissertation selected eight features which have obvious features of eigenvalues at the first, then enter these features into the C4.5 classifier to train. According to the results of training, we chose five best feature vectors finally, and then input them to the classifier and make classification on the images. At the same time, we also take into account some images which used to describe the human face in real life, this dissertation added the feature of face area percentage. Here, we related to face detection, this dissertation used a combination of AdaBoost and the Cascade approach to detect face. The experimental results show that the method not only has good results in face detection, and its speed has also been greatly improved, and meeting the system's requirements.(4) Finally, we analyzed and summarized the results of studies made by the previous, combining of accuracy in detection and real-time requirements, we constructed an erotic image filtering system framework based on skin-color detection. Among them, skin-color detection is the core of the filter system, and combining of the feature vector extracted from the mask images, and then inputting them to the classifier to achieve erotic image filtering.Experimental results show that the erotic image filters constructed by us can effectively classify the images and filter erotic images between normal images and erotic images. The precision is about 92.34% in the 5205 test image library, and the precision in erotic images is 85.74%, and in normal images is 93.24%.Although the accuracy in detection has improved, this filter system also needs to be further improved and perfected. For example, we introduced Gabor function to detect skin from the point of texture. Although the accuracy in detection improved to a certain extent, it affects detection speed of the whole system. Also, this dissertation is not comprehensive enough to consider some of the images that have the problem of highlight and shadow in the skin area, this is also one of the main reasons caused false in skin-color detection in some images. In addition, this dissertation extracted features only in the skin-color mask images, not considered the special sensitive parts of the body and other aspects of the features. These are the follow-up studies need to be improved and perfected.