基于多样本扩充的卷积神经网络人脸识别算法

Neural Network Based on Multiple Images for Face Recognition

为了更好地提取人脸特征,提高人脸识别率,提出一种基于多样本扩充的卷积神经网络CNN(Convolutional Neural Network)人脸识别算法。CNN网络能够自动提取图像深度特征,但是面对有限的人脸样本数,如何提取足够的人脸特征是人脸识别所要面对的重要问题。论文提出的新算法首先利用人脸的镜面性生成镜面图像,分别对同一类中任意的两个原始样本与镜像样本,取它们的平滑中值样本构造新的虚拟样本,将新生成的镜像样本与两类平滑中值样本作为新的训练样本集输入CNN网络得到更新的权值,然后通过目的训练样本集训练CNN提取更多隐藏的人脸图像特征,最后使用支持向量机SVM(Support Vector Machine)特征提取后进行分类。通过CNN网络和SVM能够提取更多有效的人脸特征,通过实验证明,该算法在人脸库上取得了较高的人脸识别率,具有较好的实际应用效果。

In order to better extract the facial features and improve the accuracy of face recognition,a new convolutional neural network is propose based on multiple images for face recognition.One of the advantages of convolutional neural network for face recognition is the ability to extract image features automatically.As the number of face samples is limited,it can not meet the need of face recognition in real world application.It is important to extract enough facial features.The new method firstly uses the mirror feature of the face image to generate new samples,and then uses the arbitrary two new samples and original samples respectively to synthesize the smoothing median virtual samples,these new virtual samples are introduced as auxiliary training dataset,and updated weights are gotten.Then these weights are used to update the layers of CNN,through CNN with target training dataset,more hidden facial image features are extracted.And the CNN network and SVM can be used to extract facial features more effectively.Experimental results on face databases show that the proposed algorithm can be applied to achieve high face recognition rates,it has strong ability of face description.