基于并行卷积神经网络的人脸关键点定位方法研究*

为解决人脸关键点定位受到姿态，光线，表情以及遮盖问题的影响而使得定位效果可靠性不佳的问题，提出了一种基于并行卷积神经网络的人脸关键点定位方法。文中利用卷积神经网络对局部细节特征提取以及深度学习特点，设计实现了一种并行卷积神经网络，该网络把人脸图像，上半人脸以及下半人脸分别送入到相同结构的卷积网络进行训练学习，通过对图像进行局部卷积以及下采样，提取人脸关键点附近的细节特征，并对三级并行网络定位结果进行加权合成，实现人脸特征点定位。在LWF人脸库上定位实验结果表明该方法在准确性以及可靠性都得到很大程度提升，能实现对人脸关键点的鲁棒准确估计。

Facial Key Points Location Based on Parallel Convolution Neural Network

Facial key points localization suffered from face pose diversification, illumination, expressions and collusion, the location result is unreliable.To solve these problem ,we proposed a facial key points location method which base on parallel convolution neural network. Convolution neural network can preserve the structural feature of image texture, which had widely used in face detection and OCR recognition successfully. Thus, we use the merit of CNN, designed and fulfilled a parallel CNN. Firstly, we use the entire face, up half face and down half face to train the CNN, the CNN adopts local convolution and down sampling to extract the features around the face key points, and then locating the key points. At last layer, we use the average result to fine the final result. The experiment on LWF face database showed that our approach outperforms other methods both in detection accuracy and reliability, which can locating the key points efficiently.