【摘要】 虽然人脸识别技术已经取得了很大的成就,但是如何提高不同姿态、光照、表情识别的准确性也面临着很大的挑战。为了进一步提高精度,神经网络模型被设计得越来越复杂,这也直接导致在反向传播时会出现梯度消失等现象。为了缓解这些问题,本文提出了一种复杂条件下参数较少的双通道卷积残差网络模型。该模型由多个输入通道组成,共同学习输入图像的不同特征。将原始图像作为第一通道的输入,然后利用Sobel算子提取原始图像的一阶导数特征,并将其作为第二通道的输入,来自两个通道的人脸特征信息经过融合后送入到一个残差模块,经过平均池层,最后用于识别。该网络模型结构简单、参数少、速度快、准确性高。模型在FERET、AR和FEI数据集上进行了训练和测试。实验结果表明,本文的方法优于当前的一些先进方法。更多还原

【Abstract】 Although face recognition technology has made great achievements,how to improve the accuracy of different pose,illumination and expression recognition is also facing great challenges.In order to further improve the accuracy,the neural network model is designed more and more complex,which directly leads to the phenomenon of gradient disappear in the back-propagation.In order to solve these problems,this paper proposes a two channel convolution residual network model with fewer parameters under complex conditions.The model consists of multiple input channels,which learn different features of input image together.The original image is taken as the input of the first channel,and then the first derivative feature of the original image is extracted by Sobel operator,which is taken as the input of the second channel.After fusion,the face feature information from the two channels converges in a residual module,then passes through the average pool layer,and finally is used for recognition.The network model has the advantages of simple structure,less parameters,fast speed and high accuracy.The model is trained and tested on FERET,AR and Fei datasets.The experimental results show that this method is superior to some advanced methods.更多还原