The Research On Network Pornographic Image Filtering Based On Erotogenic-part Detection

Building up content-based image filtering on the Internet is a crucial issue because of the recent explosion in the amount of online images .But the usual filtering methods are only based on the IP address and the web text content .And even if combined with a content-based image detection, the precision was always unsatisfactory when popularly employing the simple static skin-color detection algorithm. In this paper, a method for detecting skin color which is based on human face and body positional information have been proposed. And based on the body trunk search and skin-color detection, an erotogenic-part recognition algorithm is introduced. Eventually, a network pornography image filtering system is constructed .Concretely, the main contributes of the paper are:1) We review the development of face detecting technology .In the view of veracity and velocity, we select Viola's rapid face detecting algorithm which is based on the Integral Image and the AdaBoost training .The method make use of the Integral Image to calculate the rectangle features which will compose a strong classifier after the AdaBoost training .Because the rapid calculating of the integral image ,the simple-to-complex strategy and the cascade structure of the classifier ,the classifying speed increase greatly .We implement the face detecting with the Intel corporation's OpenCV image processing library .In order to successfully detect rotated faces ,each input image is passed to face detector three times .The first time is the original image ,the next two times are its rotated variants with rotation angle of 45 degrees anticlockwise and clockwise respectively .On the base of rotated face detection we calculate the average and the variance of quantum in the YCrCb color space as the face skin features .Then we employ the thinking of ellipse fitting to establish a human body model with a series of mathematic formulas .As a result, we achieve to getting the body approximate location .Because there is large naked skin area in erotic image ,the body here means the largest part of exposed skin area in this algorithm.2) During the summary of classic skin algorithm ,we analyze the character of the skin as well as the color space .Comparison has been made among several skin color models .The skin color models include Statistical Color Model; Gaussian Mixture Model; Chroma Space Model .The classic skin algorithm also use the texture detecting for the filtration .After the analysis and comparison of the texture usual algorithm ,our system introduce the one-rank-gray stat considering the result and speed .In the classic skin detection we only use the color information and ignore the importance of the pixel position information .To our common sense ,even the same color in different distribution maybe present different object .And the skin pixel appear more in the center area and around the object than other area .Aiming at the deficiency of the traditional skin color algorithm, we propose the dynamic threshold method based on the pixel positional information using the body approximate locating algorithm. Firstly we define relative distance to measure the pixel position .Here the relative distance means the minimum of the distance from the pixel to the center of the face ellipse and the body ellipse .Secondly we take advantage of the Bayes theory to establish the relation between the pixel distance and the skin probability .The skin probability of the pixel will be larger when its relative distance is more little .Then the skin threshold can be dynamically determined according to the skin probability and the face skin feature. And based on the approximate trunk location and dynamic skin-color detection, we ascertain a geometric region containing trunk by region growing and extracting region of interesting. And we adjust the trunk boundary more accurately in it by virtue of skin-color boundary.3) The implement of erotogenic-part detector has achieved on the basis of trunk location. It is constituted by two sub detector: breast feature detector and pubes feature detector. In the breast feature recognition, we firstly transform and strengthen original image to make the breast feature more obvious. Then recognize the candidate feature to breast one by virtue of some attributes of the breast feature's own, such as ellipse-like shape, the amount of edge pixels, the area of skin pixels surround it. And in the pubes recognition, several pubes filtering windows like"凹"are proposed which are similar to Haar-Like feature. And the whole image is scanned with these windows in different size and at different orientations to search and recognize the pubes feature according with the conditions given. These conditions is concerned with the ratio of the skin area, the dark area and the edge pixels'amount within the window,. And a rapid calculation of pixel summation within a arbitrary rectangle has achieved by virtue of two integral images,i.e. Summed Area Table (SAT) and Rotated Summed Area Table (RSAT) which are presented by Viola.4) We construct a multilayer pornography image filtering system by integrating traditional method and our new approach. At first, we need to decide which kind of skin-color features should be chosen for classification according to the result of face detection. If a face is recognized, the new algorithm based on dynamic skin detection, trunk location and erotogenic-part recognition is used. Otherwise, we merely extract the skin features by employing traditional method. We chose 993 images of normally dressed person, 266 images of much bare whereas not erotic person and 193 pornographic images all crawled from Internet. Then the ability of our classifier is seriously tested, and a thorough comparison was conducted between the previous skin-color detection algorithm and our new approach. As shown, the new method diminishes the false positive rate by about 20% to 37.97%.And it raise the detection rate by about 15% to 80.31%. Its total detection rate is 81.70% in all 1259 benign images.5) At last, we introduce the application of pornography image filtering system in the IE browser. First the event mechanism of IE browser and the function of its assistor object are explained. Then we introduce the approach to capturing specific event using ATL. Eventually, we present the framework of our filter plug-in and the work flow of main models.Results on extensive experiments indicate that, comparing with many previous methods, this system efficiently improve the precision and moreover, decline the false positive rate dramatically. Especially, in the case of two kinds of images below: a person dressing less but not naked; and a certain bare part of the person occupies a large area of the image due to the visual angle whereas the image is still not pornographic (e.g. the feature of hand or back), our method is expected to be more effective.