人脸识别中AlexNet网络设计和改进方法研究

Alexnet Network Design and Improvement Methods in Face Recognition

针对传统卷积神经网络模型在静态环境下人脸识别的过拟合问题,AlexNet网络的各隐层通过应用"Dropout"方法得以解决。但这个网络较复杂,计算量大,训练集通过验证集测试的准确率提升的太慢,人脸数据的损失值曲线和识别率曲线都存在振荡问题。因此本文结合Caffe深度学习框架中的AlexNet网络结构进行改进,利用梯度下降法对批量的图像数据进行特征提取器和分类器的训练。在原网络的基础上,删除一个全连接层,同时放弃使用LRN层,根据VGG网络的结构,用7×7和5×5的两个小卷积核替代原来的11×11的大卷积核,来弥补去掉的全连接层和LRN层,这样网络参数减小从而加快计算速度,人脸数据的损失值和识别率的振荡程度减小,从而达到平稳。

Aiming at the problem of over-fitting of face recognition in the static environment by the traditional convolutional neural network model,the hidden layers of AlexNet network can be solved by applying the"Dropout"method.However,this network is more complicated and computationally intensive,and the accuracy of the training set is improved too slowly by validation set test,and there are oscillation problems in the loss value curve and the recognition rate curve of face data.Therefore,the improvement is done by combining AlexNet network structure in the Caffe deep learning framework,while the gradient descent method is used to train the feature extractor and classifier of the batch image data.On the basis of original network,a fully-connected layer is deleted and the use of LRN layer abandoned.According to the structure of VGG network,the original 11×11 large convolution kernel is replaced by two small convolution kernels of 7×7 and 5×5,thus to make up for the removed fully-connected layer and LRN layer.In this way,the network parameters are reduced,the calculation speeded up.And the loss value of face data and the oscillation degree of recognition rate are reduced,thus achieving the stability.