改进的ResNeXt50神经网络面部表情识别方法

Facial Expression Recognition Method Based on Improved ResNeXt50 Neural Network

为了减少现有基于通用架构的深度学习表情识别方法在卷积过程中丢失重要信息的现象,文中提出一种改进的ResNeXt50网络(命名为AC-SP-ResNeXt50),以ResNeXt50为基础架构,增加一个多尺度特征提取层,以不同尺寸的卷积核对原始图像进行特征提取,提取到更多纹理信息;同时以SoftPool作为网络的池化层,保留了更多特征信息;使用非对称卷积模块构成残差结构,强化了特征提取;文中通过消融实验和对比实验,验证文中方法在减少特征信息损失,提升面部表情识别率方面的有效性;通过识别随机选取的表情图像,评估了文中模型的泛化能力。实验结果表明:多尺度特征融合、SoftPool、非对称卷积对面部表情识别结果均有提升作用。文中方法在CK+数据集上的识别率可达到98.93%,在Jaffe数据集上可达到97.62%,与卷积神经网络与支持向量机相结合的方法(CNN+SVM)、注意力与空间注意力机制分离方法(CA-ST-DSC)、全局分支和局部分支结合的方法(GL-DCNN)、基于深度可分离卷积的识别方法(DSC-FER)等现有面部表情识别方法进行对比,文中方法在特征提取方面更具优势,识别结果更佳,对现实中的表情图像识别能力也较为出色,模型泛化能力较好。

The existing deep learning expression recognition methods based on general architecture tend to lose some important information during the convolution process.To solve this problem,an improved ResNeXt50 network(named AC-SP-ResNeXt50)was proposed,with ResNeXt50 as the basic architecture and a multi scale feature extraction layer added.Convolution kernels of different sizes were used to extract features from the original images and more texture information were obtained.SoftPool was used as the pooling layer of the network to retain more feature information.Asymmetric convolution modules were used to form the residual structure,which strengthened the feature extraction.Ablation experiments and comparison experiments showed that the proposed method is effective in reducing the loss of feature information and improving the recognition rate of facial expressions.The generalization ability of the model was evaluated by identifying expression images selected randomly.The experimental results show that the multi scale feature fusion,SoftPool and asymmetric convolution all can improve the recognition results.The recognition rate of the method in this paper can reach 98.93%on the CK+dataset and 97.62%on the Jaffe dataset.Compared to the existing facial expression recognition methods,such as“the method which combines convolutional neural network and support vector machine(CNN+SVM)”,“Attention and Spatial Attention Mechanism Separation Method(CA-ST-DSC)”,“A combination of global branch and local branch(GL-DCNN)”,“the recognition method based on depthwise separable convolution(DSC-FER)”,etc,better results are achieved by the new method,which has a better ability to recognize facial expressions in reality and better model generalization ability.