Research On Face Detection Based On Deep Learning

Face detection is a computer technology that determines the locations and sizes of human faces in an arbitrary digital image. It is a challenging task for computers because of variability in lighting conditions, pose, facial expression, noises and occlusion. Deep learning is a sub-field within machine learning that is based on algorithms for learning multiple levels of representation in order to model co mplex relationships among data. It is widely used in sigal and information processing. We detect faces in static gray images by utilizing deep neural network classifier.Face detection is a task of classification of two patterns. The main work of face detection is to train a classifier that separates faces and non-faces accurately. The two steps of face detection are training of classifier and face localization. In the first step, we trained four kinds of classifiers, Deep Belief Network, Stacked De-noising Autoencoder, Deep Neural Network without pre-training and Multilayer Perceptron without pre-traing, respectively. Deep Belief Network and Stacked De-noising Autoencoder have the smallest classification error on both training set and testing set, which implies that the features extracted in unsupvised pre-trainging are helpful to improve the power of model’s discrimination. We choose the Deep Belief Network which has lowest generalization error as a classifier. In the face localization step, we proposed a face detection method based on the trained Deep Belief Network classifier. First, we search face candidates in the pyramid maps of input image. Then, we map these face candidates to the original input image. Thus, we get face candidates which have certain size and location in the input image. To eliminate overlap regions, our own rules are used to merge them.Finally, the proposed detection method is tested on the set of 130 upright faces in CMU database with detection rate 91% and false positives 100. Frontal faces within rotation range [-45, 45 ] and profile faces within yaw range [-45, 45 ] in the complex background images can be effectively detected.