基于HOG与卷积神经网络的人脸识别研究

A Face Recognition Method Based on HOG Feature and CNN

为了获取更加全面的整体与局部人脸特征,得到更高的人脸识别率,提出一种基于方向梯度直方图(HOG)特征与卷积神经网络的人脸识别新方法。该方法首先提取人脸图像的HOG特征,然后将HOG特征图像作为卷积网络的输入数据进行训练,改进网络结构,在全连接层之后采用Softmax loss和center loss两个损失函数进行监督,最后在训练得到的网络模型上对人脸图像进行识别操作。实验结果表明,该方法在ORL人脸集上的识别率达到97.5%,相比于其它人脸识别算法具有一定优越性。

In order to obtain a more comprehensive global and local facial features and a more accurate face recognition rate,this paper proposes a new method based on the combination of direction gradient histogram(HOG)features and convolutional neural networks.The method first extracts the HOG feature of the face image,and then trains the HOG feature images as input data of the convolution network,changing the network structure,and adopts Softmax loss and center loss after the fully connected layer.Finally,the face images are identified on the trained network model.A 97.5% recognition rate is obtained on the ORL face set,which is superior to other face recognition algorithms and has certain advantages.