基于卷积神经网络的表情不变三维人脸识别

Expression invariant 3D face recognition using convolutional neural networks

针对三维人脸识别中的表情问题,提出一种基于卷积神经网络的三维人脸识别方法。根据人脸先验知识,构建基于测地线距离的三维人脸特征点模型;利用该模型,提取输入三维人脸的局域Gabor特征和测地线距离特征,进而获得表情不变的人脸表述;将上述特征输入类Lenet-5卷积神经网络,获得最终的识别结果。在Facewarehouse三维人脸数据库上的实验结果表明,该方法的正确识别率达到97.60%,优于几种经典三维人脸识别方法,对表情变化均有较强的稳健性。

Facial deformation is an urgent problem to be solved in 3Dface recognition.This paper presented a 3Dface recognition method based on convolutional neural networks,which utilized the deep learning to realize the automatic feature extraction and classification,which is robust to expression variance.Feature points and their topological structure were determined on the average face according to priori knowledge.By 2D Gabor wavelets,robust neighborhood feature of each point,combined with geodesic distances were extracted.These features above were took as the input of convolutional neural network,which used the model similar to Lenet-5.After the train,utilize this model to finish recognition.Experiments on Facewarehouse 3Dface database demonstrate that our method can improve recognition accuracy（97.60%）.To sum up,the proposed method could improve the recognition rate under facial deformation,and deep learning provides a path for 3Dface recognition.