【摘要】 针对融合卷积神经网络学习到的低层次特征与高层次特征进行表情识别时参数过多的问题,提出面向表情识别任务的双通道卷积卷积神经网络。将池化层池化得到的特征图分为上下两路进行卷积,上路特征图采用1×1卷积核进行卷积得到低层次特征值,下路特征图输入到下一卷积层中学习高层次特征,将高层次特征与各层的低层次特征相融合后输入分类器进行分类。多个表情数据集实验结果表明,所提方法在保证识别精度的前提下有效减少了特征融合后的参数量。更多还原

【Abstract】 In the task of facial expression recognition,to make full use of the low-level features and high-level features which learned using convolutional neural network,and to reduce the parameter dimension,a double-channel convolution convolutional neural network was proposed.The feature maps outputted by pooling layer was divided into two channel for convoluting,the upper channel was convolved with 1×1 convolution kernel,low-level feature values were got,the lower channel was inputted into next convolution layer to learn high-level features.Low-level features and high-level features were fused and inputted into classifier.Experimental results on two datasets show that the proposed method can effectively reduce the amount of parameters.更多还原