Research On Content-based WEB Image Filter Technology

With the development of Internet, people can contact all kinds of material on the world-wide-web. Pornography content is harmful for under-age net surfer, this is a serious social problem. How to prevent children from accessing pornographic content on web is a growing problem parents concern about. Besides legislation way, it calls for effective techniques to keep children from accessing these objectionable content. There are three main content-based WEB filter method, URL-based filter, text-based filter and image-based filter. The first method has to maintain lists of URL that need to be updated frequently manually, the second method may failed to work when there is no pornographic text or text is displayed in picture mode. Since adult image is main characteristic of adult web site, image-based filtering is a promising way. This paper proposes an effective system to detect adult image, it consists of skin detection, face detection, features extraction and image classification.Base on the fact that there is a strong correlation between image with large patches of skin and adult image. skin detection is of the significant importance in the image-based filter system. It can keep skin patches and rule out trivial background, which brings convenience to subsequential process. Traditional skin detection method only use pixel-level color information which may mistakenly classified things resemble human skin in color as human skin. An improved skin detection method integrates color, texture and space information is proposed. To reduce false acceptance rate, after SPM color filter, a texture filter was constructed based on texture features extracted form Gabor wavelet transform. To increase true acceptance rate, pixels pass both color and texture filter were used as seeds to implement diffusion process. After skin detection, morphological filter is applied to reduce noise, holes in skin patch and false objects in background.Quite a number of WEB images are the kind of image that face occupies most space in the image(close-up face image is typical example). These kinds of images are not adult images, so they should be ruled out at the early stage, which will eliminate error maybe made by subsequential process. Take speed and accuracy into account, we adopt AdaBoost face detection algorithm. Since AdaBoost face detection algorithm only exploit grey information, it may mistake thing resemble face in grey attribute as human face. To remedy the mistake, we utilize the output of skin filter to make color verification, calculate Euler number to make geometric verification. WEBimages were formed under unconstrained environment, some image may exist serious color distortion due to abnormal color temperature. After obtaining face object in image, we use skin color in face as reference color to perform white balance, then re-implement skin detection to improve detection accuracy.Though most nude images contain large patches of skin, relying solely on color information to identify nude image is insufficient. To recognize nude image reliably, machine-learning technique should be introduced. Therefore, a set of features that can capture the main characteristics of nude image must be extracted first. We extract skin-related features from the output of the skin detector, extract color features from color moments, chromaticity moments and color correlogram, extract texture features by use of wavelet packet analysis, extract shape features from Hu invariant moment and its improved form that computed based only on the shape boundary. These features compose a joint feature vector fed to classifier.In classification stage, our task is to find the decision rule that can optimally separates objectionable images from benign images. We take this task as a two-class pattern classification problem. SVM was originally designed for binary classification problems, it pursues structure risk minimum instead of empirical risk minimum that can guarantee its generalization ability. So we adopt SVM as our classifier. The performance of SVM is fixed once its train process is finished, which is inadequate to classify infinite new unconstrained images. Since it is impractical to collect large collections of image of all types, we combine the advantage of SVM active learning, SVM incremental learning and SVM transductive learning, incorporate a long-term learning scheme into our system. This flexible scheme may automatically adjust and improve performance of SVM according to newly encountered test data.