Facial Analysis Via Deep Learning Methods

Long with the development of society and computer equipments, facial analysis becomes a more and more important task in computer vision with many applications. This thesis is focused on thress aspects: heavily occluded human faces localization, kinship verification and smile detection.Localizing heavily occluded human faces is a challenging problem in facial detection. Previous methods mainly employ sliding windows by determining whether windows include human faces. In this paper, we provide a novel segmentation-based perspective for heavily occluded face localization with deep convolutional neural networks(CNN). Our model takes an image as input. After several convolutional layers, fully-connected layers and a softmax classifier, we can predict the labels of pixels, which is the key to localize heavily occluded human faces. Finally, we search a minimal rectangle to localize the human face. Our detector does not need time-consuming sliding window. Besides, we use a single model to localize faces to further alleviate computational complexity. Experimental results show that our proposed method achieves encouraging accurate with only 4 milliseconds per image.Kinship verification from facial images is an interesting and challenging problem. However, manual features cannot well discover information implied in facial images for kinship verification, and thus even current best algorithms are not satisfying. In this paper, we propose to extract high-level features for kinship verification based on deep convolutional neural networks. Our method is end-to-end, without complex pre-processing often used in traditional methods. The high-level features are produced from the neuron activations of the last hidden layer, and then fed into a soft-max classifier to verify the kinship of two persons. Considering the importance of facial key-points, we also extract key-points-based features for kinship verification. Experimental results demonstrate that our proposed approach is very effective even with limited training samples, largely outperforming the state-of-the-art methods. On two most widely used kinship databases, our method achieves improvements compared with the previous best one.Smile detection from facial images is a specialized task in facial expression analysis with many potential applications. In this paper, we propose to extract high-level features by a well-designed deep convolutional networks(CNN). A key contribution of this work is that we use both recognition and verification signals as supervision to learn expression features, which is helpful to reduce same-expression variations and enlarge different-expression differences. High-level features are taken from the last hidden layer neuron activations of deep CNN, and fed into a soft-max classifier to estimate. Experimental results show that our proposed method is very effective, which outperforms the state-of-the-art methods.