| With the widespread use of cameras in recent years, the study of skin region segmentation in image is becoming more and more popular. Techniques of skin detection are used more and more in machine vision systems. Skin detection is currently mainly used in face detection and recognition, human-computer interaction, image and video retrieval, Filtering pornographic pictures, etc. A relatively wide range of applications are also in computer graphics, the diagnosis of skin diseases, the effect on the analysis and the test of cosmetics lies, etc.Video during the shoot will be affected by the angle and color of light and so on, which will occur the color shift; At the same time the surface of objects affected by specular reflection or interface reflection and the images often cause high light or shadow. In the video frame, the high light areas usually corresponds to the bright areas, shaded areas corresponds to low brightness areas; Different racees will appear in the video; The color style of the whole video showing (black and white, not bright, more bright, very bright, etc.) is different and so on. These factors will make an impact on adaptive skin color modeling on accurate. Therefore, first we should be prior to pre-processing video before adaptive skin color modeling.A method for skin modeling is provideing the threshold range of color spaces, in this paper we selected YCbCr space, first we converse the video frame from RGB space to YCbCr space, and comparing corresponding Y value of each pixel in image, if the range of the brightness is in skin model, and the corresponding chroma (Cb, Cr) value of each pixel is within the range of image corresponding to the pixel,and we assigned a1, otherwise assigned O,which will get a binary image. If the background of the video frame is relatively simple, The method can well detect skin regions, but the background of a video frame is complex, accurate detection of skin color is a certain difficulty.We use the methed of region growing to detecte skin region, first we should find a seed point, and then merger the pixels into a region around the seed point.We need find a suitable threshold T in the merger, if the threshold is too small, skin is not easily detected, whereas, T value is too large, and the skin region will included the similar background. These two methods above can not meet the requirements of skin modeled in complex background image, this paper studies an adaptive skin modeling. Adaptive skin color modeling process is, first we use AdaBoost algorithm to detect human faces in video frames, and then use these detected face region color information to model. In the study, in order to achieve better color modeling results, first we appropriately reduced the detected face area by the AdaBoost algorithm; then use multi-color space information fusion technology, by selecting more than one color space components (Y, Cb, Cr, H, S, I, Y, U, V), calculate the SPM (color probability map) of these components in the test image to be selected,get the final test of the skin area after "and" operational integration. The algorithm does not consider the color shift, color, style and race and other factors, is an adaptive modeling process.This paper focuses on the process of skin model in sex video detection and solve the problem of adaptive skin color modeling. The main contents are as follows:1. We use the cumulative Hsitogram of S (saturation) component to complete video frame style classification, and use "GrayWorld" approach to correcte the cast of the video frame images for color. The style of video frame is divided into black and white color, not bright, more bright, very bright and so on. And correct the image of color cast to complete the video pre-processing for adaptive color modeling.2. We use AdaBoost algorithm to detecte faces. AdaBoost algorithm use a large number of classification ability of weak classifiers together to form a very strong classification ability of classifier. Theory that, as long as the classification capability of each weak classifier is a little better, and when the number of weak classifier tends to infinity, the error rate of strong classifier will tend to0. We use the detector to detecte faces of635individuals,51missed,19false alarms, detection rate of88.9%, Use AdaBoost algorithm to face detection, Complete the face detection complex of context.3. We use multi-color space and skin probability map (SPM) for adaptive skin modeling. Then use the multi-color space information fusion technology, namely, by selecting a number of multiple color space color components, calculate the each SPM (skin probability map) of these components in the detected image, get the final test of the skin area after "and" operational integration. Light, color and ethnic origin and so on are not considered in this algorithm,the method is an adaptive modeling process. Currently we selected HCbCr as color space, can not be able to select color components, which is the next work we need to improve and explore. To further improve the color accuracy of modeling. |
PDF Full Text Request