Facial Landmark Detection Research In The Wild

Face plays an important role in visual communication.Human beings can automatically extract many non-verbal information by observing faces,such as human identity,intention and emotion.In all face analysis tasks in computer vision,facial landmark detection is usually a necessary and key step for extracting the facial information automatically.Therefore,a simple and effective facial landmark detection algorithm is indispensable for the task of face analysis.In order to solve the problem that the current facial landmark detection algorithm based on convolutional neural network is too complex,an effective deep learning framework based on generative adversarial network is proposed.The network consists of a generating network and two discriminant networks.The generating network consists of an encoder and two decoders.The encoder predicts the coordinates of the key points of the face,and the decoder generates the geometric feature maps of the face interior and contour according to the predicted key points.Discriminant networks discriminate whether a given geometric feature map is generated by ground truth or the generated network,Discriminate whether the input geometric feature map is generated by real key points or by prediction of the generating network,and return the residual to optimize the generated network.Finally,the generating network can predict the coordinates of key points close to the real value through continuous game training.The proposed network is an end-to-end training structure.In the detection stage,only the encoder is used as the facial landmark detector.This paper also proposes an on-line hard sample mining algorithm to improve the generalization ability of the network model for unknown targets.Improving the diversity of samples through random data enhancement,a strategy of selecting hard samples is designed to select the best training samples from the enhanced data set.Comparing with other facial landmark detection algorithms in the wild,it is found that this method has the expected performance on many data sets.Besides the model is more simple,and the number of parameters is smaller.However,the detection performance of the algorithm in extreme environments such as extremely low illumination and occlusion still needs to be improved.