| The facial expression plays an important role in our social interactions as it conveys rich sources of information.The thesis concerns using computer vision methodologies to analyse the micro-expressions of one of the facial expressions that can hardly be perceived by naked eyes.Micro-expressions are rapid,involuntary facial expressions that belong to an essential non-verbal behavior which reveal emotions people do not intend to show.At the same time,because micro-expressions are difficult to be perceived by the producers,thus automatic micro-expressions analysis has a wide range of applications in the national security and computer aided diagnosis.However,the images captured from surveillance video easily suffer from the low-quality problem,which leads to great challenges for conventional microexpression recognition algorithms in real-world applications.For addressing the above problems,this thesis conduct a comprehensive study about the micro-expression recognition problem under low-resolution cases,and proposes:(1)On the basis of the optimization of the active shape model facial calibration algorithm,facial hallucination method is used to super-resolution reconstruct micro-expression image sequences under low-resolution cases from two aspects of texture-based and structure-based.Fast LBP-TOP feature extractor is used to extract the spatio-temporal information of reconstructed image sequence,and linear support vector machine is used to classify the above information to obtain the final accuracy of micro-expression recognition.(2)Aiming at the problems of lengthy process and low accuracy of micro-expressions recognition by the traditional machine learning method under low-resolution cases.Based on the fusion and augmentation processing of existing databases,the pseudo-3D network structure based on residual neural network is proposed to realize the end-to-end automatic micro-expression recognition task.(3)The thesis builds the first automatic micro-expression visual recognition system under low-resolution cases(LMERS).Visualization is used to promote the study of microexpression recognition under low-resolution cases.The thesis proposes method based on traditional machine learning is experimented on two degraded databases and good recognition results are obtained,which breaks the situation that existing algorithms can not recognize low-resolution micro-expressions.And the method based on deep learning achieves higher accuracy on enhanced databases than traditional machine learning. |
