卷积神经网络在人脸识别上的研究

摘要

卷积神经网络在人脸识别领域里已经有非常好的结果,在LFW(Labeled Faces in the Wild)数据集上的认证识别率已经达到99%以上,超出人的识别率.然而这些网络需要训练庞大的参数或者需要巨大的数据集.通过构造一个比较小的网络,极大地减少了训练参数,并且在有限的训练集上得到了比较好的识别率.为了深入了解卷积神经网络,做了很多实验.例如,比较了在激活层前和激活层后加入正则化对网络训练的影响,同时比较了提取不同的特征维度对实验准确率的影响,另外也比较了欧式距离和余弦距离对实验的影响.通过实验结果我们发现正则化在激活层之前的网络识别率更高,使用余弦距离比使用欧氏距离会有更高的识别率,特征维度越大,人脸识别率越高.%Convolution Neural Network (CNN) has achieved promising results in face recognition recently .The verification recognition accuracy of some very deep neural network has reached 99%, exceeding person's performance.However, these networks need to train a vast number of parameters or huge dataset.In this paper, a smaller network, greatly decreasing train parameters and obtaining better recognition rates within limited dataset was put forward .In order to deeper understand convolution neural network , we made many experiments had been made .The effects of normalization before and after the active layer on network training were compared .Meanwhile, the effects of different feature dimensionality on recognition accuracy were compared , and the effects of Euclid distance and cosine distance on recognition accuracy were compared .From the results of experiments , the network which normalization was before the active layer gets higher accuracy .The way of using cosine distance was better than that using Euclid distance , and the feature dimensionality and accuracy were higher .