Micro-expression Recognition Based on ME-Xception Convolutional Neural Network

‍ ‍‍ ‍Micro-expression has the problems of small facial muscle motion range and few dataset samples， which will make it difficult for neural network to capture effective features in the learning process and improve accuracy. Therefore， a micro-expression recognition method based on improved Mini-Xception convolutional neural network is proposed. Firstly， in the preprocessing stage， the magnification is calculated according to the cosine similarity， and the micro-expression is amplified adaptively. Then， the Mini-Xception model is improved. The specific operation is to add projection layer on both sides of the input layer to reorganize the input characteristics， and add the channel attention mechanism to the circular module composed of deep separable convolution layer and batch normalization layer to construct the ME-Xception model. Finally， the ME-Xception model is applied to the task of micro-expression recognition. Experiments are carried out on CASME Ⅱ， SAMM and SMIC datasets. The results show that this algorithm effectively improves the recognition accuracy and can obtain better recognition performance compared with other mainstream algorithms.