Face detection is to look for and locate face patches in given digital images. As typical non-rigid object detection, face detection has broad application prospects, such as content-based image and video retrieval, digital image processing, intelligent human computer interaction, biometrics, and so on. Its performance has great effect on computer vision, pattern recognition and even the entire compute science.In order to detect human faces among color images which have complex backgrounds, this thesis is devoted to the color-based face detection and support vector machine (SVM) learning method.Skin-color segmentation is based on that although different people have different skin color, but the major difference lies largely in their intensity rather than their chrominance. Many color spaces have been used in segmenting skin regions in color images, such as RGB, normalized rgb, HSV, HSI, YCbCr, and so on. Literature survey shows that the YCbCr color space is one of the successful color spaces in segmenting the skin color accurately, mainly because the chrominance components are almost independent of luminance component in the space. According to the statistical data of skin samples which was manual collected, we built a simple Gaussian skin-color model in the nonlinear YCbCr color space. Compared with the traditional projection and simple Gaussian models built in YCbCr color space, the model this thesis established can better enhance color contrast between the skin regional and non-skin regional. This makes the binary image of face candidates which gained by skin-color segmentation better reflect the detailed information, such as the eyes, eyebrows and mouth.Skin-color model is not dependent on the details of facial features, so it can be used in the image with rotated face and complex expression. The model is real-time and relatively stable. However, it is inevitable that the face candidates comprisesome non-face skin regions and skin-similar regions, such as arms, hands, complex backgrounds. Then the accuracy of skin-color segmentation will be lowed.SVM is a popular binary classification, which aims to minimize an upper bound on the expected generalization error and structure risk. The idea of SVM is to project the input data into a high-dimensional implicit feature space at first with a kernel mapping, and then to look for an optimal classification hyper-plane in the feature space. SVM has been widely used in pattern recognition due to its good classification properties. Although SVM has solid theoretical foundation, it is difficult to determine the parameters which have great effect on the VCdim. Therefore, this thesis was researching in the problem of parameters. Comparing Cross -Validation with leave-one-out Validation, we adopted the 5-Cross-Validation based on the grid search strategy.The training set comes from the MIT-CBCL, it has 6000 train images, and the image size is 19*19. If training the SVM classifier directly, the High-dimensional training matrix will result in a long time. Therefore, this thesis reduces the dimensional through PCA.PCA is an important analysis method of multivariate statistical analysis. It is an orthogonal linear transformation that transforms a number of correlated variables into a number of uncorrelated variables called principal components. The first principal component accounts for as much of the variability in the data as possible, and each succeeding component accounts for as much of the remaining variability as possible. Specifically, using singular value decomposition theorem seeks the eigenvalues of training matrix, and then constitutes the face PCA eignspace with the eigenvalues which are correspond to the top 120 eigenvectors. Finally, project the train images onto the PCA eigenspace and the projection coefficients are the new training set. The training of pca-SVM takes about 1/3 time compared with training of SVM. As a face classifier the pca-SVM trained by this thesis performs well, but it difficult to meet teal-time requirements.As discussed above, A Scheme of face detection based on skin-colorsegmentation and pca-SVM classifier is proposed in this thesis, and a multi-scale Face Detection System is constituted.First, color model is performed to segment the candidate image to pre-exclude non-skin regions. Secondly, through connectivity analysis, apriori knowledge such as aspect ratio is adopted to exclude non-face skin regions and skin-similar regions. Thirdly, the rest candidates are compressed into pyramid image sequence with 1.2 compression ratio in order to complete multi-scale face detection. Finally, verify the candidates with pca-SVM classifier, and mark the face regions. The experiment result shows that the method in this thesis is much effective in the detection of face in complex background. |