Research On Sensitive Image Filtering Algorithms Based On Content Analysis

With the development of network technology, the flooding of network eroticism becomes more serious. The flooding of network eroticism not only badly affects the physical and mental health of teenagers, but also brings many inconveniences to people's life. How to filter erotic information off Internet is an important research topic. In conjunction with the Innovation Project "Content-Based Erotic Image Recognition and Monitoring System" which we cooperate with China Criminal Police University, we study the key technologies of erotic image filtering based on skin-color and texture. We intend to develop an effective erotic image filter at the end.The main research of this thesis includes skin-color detection algorithm, texture detection algorithm, and face detection algorithm.Based on these researches, this thesis develops two classifiers with the features are extracted from these researches; one is Binary Decision tree, and the other is support vector machine. Skin-color detection is the most important of the erotic image analysis. This thesis investigates Duan Lijuan Chroma model and HSV skin-color model and has found the deficiency of them. And based on them, the thesis optimizes the skin-color detection algorithm. Texture is an important feature which can distinguish the erotic image from others.In the erotic image filtering system, the big-head images always are recognized as erotic ones. Therefore, this thesis investigates AdaBoost fast face detection algorithm, analyzes its deficiency, and finds where can be improved. According to the characteristics of erotic image filtering system, this thesis improves AdaBoost face detection based on geometrical shapes. In order to accomplish the system, this thesis extracts many features, selects three effective features as the feature vector through experiment.The results of experiment indicate that the methods and models given in this thesis can classify the erotic images and normal images well.The final classification result can achieve Recall Rate 88.58% and Misjudgment rate 19.01%.