一种面向移动端的浅层CNN表情识别

移动端的表情识别有巨大需求，但是受算力限制，主流深度神经网络无法直接移植。为此，设计了一个浅层网络，在节约计算量的同时保证了识别率。网络中使用三组堆叠而成的卷积层，有助于增大感受野，便于更好地提取特征，这是提升识别率的关键；使用全局平均池化层，避免引入额外的全连接层，大幅降低参数量，在训练样本不足的情况下，降低模型过拟合风险。在FER-2013数据集进行训练，准确率超过现有大多数算法；在CK+数据集上进行微调，测试集上的准确率可达到0.96。将所得模型转换为Core-ML模型，结合Xcode平台在iOS端搭建了实时表情识别App，在iPhone 8 Plus上能够稳定、流畅运行，识别效果达到预期。

Mobile-Oriented Facial Expression Recognition Based on Shallow CNN

There is a huge demand for facial expression recognition on mobile terminals.However，due to the limitation of computational power，most popular deep neural networks cannot be directly transplanted.Therefore，a shallow network is designed in this paper，which can not only save the amount of calculation，but also ensure the recognition rate.The network uses three groups of stacked convolution layer，which helps to increase the receptive field and facilitate better feature extraction.This is the key to improve the recognition rate.The network also uses the global average pooling layer instead of additional full connection layer，which greatly reduces the number of parameters，and reduces the risk of over fitting in the case of insufficient training samples.The accuracy of model is higher than that of most existing algorithms on FER 2013 data set.The accuracy rate can reach 0.96 by fine tuning on CK + data set.The model is transformed into the core ML model，and a real time expression recognition app is built on the iOS side with Xcode platform.It can run stably and smoothly on the iPhone 8 plus，and the recognition effect reaches the expectation.