基于卷积神经网络多层特征融合和稀疏表示的表情识别

In the convolutional neural network, the shallow features have much more detail information of image，the deep features include more abstract information of image, to take advantage of shallow and deep features, the method of fusing shallow features and deep features is proposed. Sparse representation algorithm is adopted to calculate the correlation between convolutional layer features and label, choosing the more relevant features with labels, thus dimension reduction is achieved. SVR regression algorithm is used to train and predict facial emotion based on the Arousal-Valence continuous emotional space model. The contracs expression results show that the proposed method significantly improve facial expression recognition.