结合迁移学习与可分离三维卷积的微表情识别方法

Micro-Expression Recognition Method Combining Transfer Learning and Separable 3D Convolution

针对现有微表情自动识别方法准确率较低及微表情样本数量不足的问题,提出一种融合迁移学习技术与可分离三维卷积神经网络(S3D CNN)的微表情识别方法。通过光流法提取宏表情和微表情视频样本的光流特征帧序列,利用宏表情样本的光流特征帧序列对S3D CNN进行预训练,并采用微表情样本的光流特征帧序列微调模型参数。S3D CNN网络由二维空域卷积层及添加一维时域卷积层的可分离三维卷积层构成,比传统的三维卷积神经网络具有更好的学习能力,且减少了模型所需的训练参数和计算量。在此基础上,采用迁移学习的方式对模型进行训练,以缓解微表情样本数量过少造成的模型过拟合问题,提升模型的学习效率。实验结果表明,所提方法在CASME II微表情数据集上的识别准确率为67.58%,高于MagGA、C3DEvol等前沿的微表情识别算法。

The existing automatic micro-expression recognition methods are limited in accuracy,and suffer from inadequate micro-expression samples.To address the problem,a micro-expression recognition method that combines transfer learning and a Separable 3 D Convolutional Neural Network(S3 D CNN)is proposed.The optical flow method is used to extract the feature frame sequences of optical flow from macro-expression and micro-expression video samples.The sequence extracted from macro-expression samples is used to pre-train the S3 D CNN,and the sequence extracted from micro-expression samples is used to tune the model parameters.S3 D CNN consists of separable 3 D convolutional layers,which are composed by 2 D spatial convolutional layers and 1 D time-domain convolutional layers,so S3 D CNN can provide better learning ability than traditional 3 D CNN with fewer required parameters and calculations for model training.Furthermore,transfer learning is used to train the model,so the over-fitting problem of the model caused by inadequate micro-expression samples can be alleviated,and the learning efficiency of the model can be improved.Experimental results on the CASME II micro-expression dataset show that the recognition accuracy of the proposed method reaches 67.58%,higher than MagGA,C3 DEvol and other advanced algorithms.