Research On Facial Landmark Detection Based On Fully Convolutional Network

Facial landmark detection aims to locate the position of predefined semantic feature points on the face image.Since these landmarks provide a rich semantic description for the face,and remove the irrelevant information in the original image,they are important to face verification and other related tasks.In the research of facial landmark detection,methods based on full convolutional neural network has become the research hotspot in recent years since FCN preserves more low-level spatial information due to the fusion of shallow features and deep features,which is beneficial to improve the performance of the algorithm.However,in the unconstrained case,there are various changes in the face,such as posture,expression,scale,etc.,which brings great challenges to the facial landmark detection.This paper studies the problem of facial landmark detection under unlimited conditions based on the full convolutional neural network.In view of these problems,this paper studies the simplification of tasks,the perfection of full convolutional neural network architecture,and the enhancement of facial feature discriminability.The main contributions are as follows:1.We propose a cascaded hourglass network based on the hourglass structure.Our method adopted a cascaded structure.Data enhancement is used to generate various training data with different levels of granularity in different concepts.The network at different level learns different levels of features,and there is better coordination between multiple levels.2.From the theoretical and experimental aspects,we point out the shortcomings of the previous fully convolutional neural network for facial landmark detection,that is,the inconsistency between training and testing evaluation criteria.Considering the characteristic of this inconsistency,we propose a method to optimize the residual facial shape and cross entropy simultaneously.The proposed method has achieved best performance over the previous methods on the current public data set 300-W.3.We studied the problem of how to apply discriminative learning for facial landmark detection,and proposed a facial landmark detection algorithm based on constraint of local facial features.Considering the changes of face image under non-limiting conditions,we proposed to eliminate the influence of scale and pose on feature invariance by alignment first,and then add the similarity constraint of facial features of different face images.Our method enhance the invariance of features of facial landmark with the same semantics.The invariance of features further improve the performance of the facial landmark detection.Finally,we designed a ID face verification system based on facial landmark detection.