Research On Micro-expression Recognition Of Transfer Learning Integrated With Attention Mechanism

Micro-expression is a kind of facial expression that is spontaneous,hard to find,small amplitude of action and short duration,which is widely used in judicial investigation,medical diagnosis and other fields.However,traditional machine learning algorithms have low recognition rate.When deep learning is used for micro-expression recognition tasks,various strategies such as increasing the width,depth,and rich receptive fields of convolutional neural networks are usually chosen to improve the recognition performance.Due to the small sample size of micro-expression dataset,these strategies have limited effect on improving deep models for micro-expression recognition.In recent years,with the continuous development of research based on convolutional neural networks,transfer learning and attention mechanisms have become effective means to improve the generalization performance of deep networks,and also bring new directions for the improvement of micro-expression recognition models.This paper focuses on the transfer learning micro-expression recognition research integrating attention mechanism.The main contributions are summarized as follows:1.To tackle with the weak variation of micro-expression facial features,the deficiency of key features and the intensity in micro-expression sequences,two improved Resnet-18 transfer learning network models are proposed.Firstly,the best network model parameters were pre-trained in CK+,the ensemble A dataset(CK+and CASIA mixed datasets)and the fused B dataset(CK+,CASIA and fer2013)respectively;Secondly,in order to extract the tiny facial features of micro-expressions and reduce the loss of key features,a convolutional attention module(CBAM)and residual attention module were introduced and integrated into the first convolutional layer of the Resnet-18 network,respectively;Structural similarity was used to determine the keyframes of micro-expression sequences;Finally,the network is fine-tuned and migrated on the CASME2 and SMIC datasets of micro-expression keyframes.The experimental results show that the recognition performance of the proposed transfer learning network model has been greatly improved compared with the previous methods.2.To cope with the insufficient micro-expression dataset samples and imperfect label information,a conditional adversarial network model for micro-expression recognition is proposed.Firstly,the CK+ macro-expression dataset is used as the source domain,and the SMIC and MMEW micro-expression datasets are used as the target domain;Secondly,the Resnet-18 network is used to extract the feature information of the image;Finally,the pre-classified information of the classifier is compared with Resnet-18 The features extracted by the network are combined for transfer learning.The experimental results show that the micro-expression recognition method based on the conditional adversarial network model achieves better recognition results than the traditional transfer methods.