Face Alignment Based On Deep Learning

Face alignment,aims at locate some predefined landmark on the face given the face detection results,which is one of the most important tasks in the field of computer vision.Due to the influence of heavy occlusion,extreme blur,large pose and exaggerated expression,face alignment is still immature facing the complicated real world,robust face alignment remain as a challenging task.Convolutional Neural Networks have achieved promising performance on face alignment,mainly because its powerful nonlinear fitting ability and end to end learning.However,face alignment based on convolutional neural networks is sensitive to face box provided by face detection and not robust to occlusion and pose.Our main contributions aim to study the method which can reduce the influence of occlusion and boost the performance of face alignment.Our contributions consist of face alignment algorithm under occlusion,face calibration based on spatial transform network and adversarial learning,a fast face calibration based on convolutional neural network.In order to solve face under occlusion,we propose a novel face alignment algorithm based self-adaptive scoring mechanism and face shape reconstruction.Existing face alignment algorithms only predict facial landmarks,ignore the difficulties of locating landmarks.While predicting the location of landmarks,our algorithm can also judge the accuracy of landmarks.Furthermore,a face shape reconstruction algorithm is proposed to infer the exact locations of the landmarks under heavy occlusion.Experiment results on 300W,COFW and WFLW dataset show the effectiveness of the proposed method.Face detection is the prerequisite of Face alignment,due to the influence of heavy occlusion and extreme blur,face detection is too weak to provide a best-suited box in some challenging cases,which may bring performance degeneration to face alignment.To solve the issue,we propose a face calibration algorithm based on spatial transform network and adversarial learning,which can precisely crop the face region,eliminate the rotation of the face image and improve the performance of subsequent face alignment algorithms.Furthermore,we propose a fast face calibration based on convolutional neural network,different from the existing methods predict landmarks and then calculating the affine transformation coefficient,our method directly predict the affine transformation coefficient and face calibration.