Micro-expression Recognition Based On Dynamic Image

Similar to traditional macro-expression, micro-expression contains several types of basic emotion, accurate recognition of micro-expression can effectively grasp the change in mood, and then provide a basis for psychological research, behavior prediction and so on. Different from macro-expression, micro-expression induced mechanism is complex and subtle with short, accompanied by a weak and short process from beginning to end, the previous analysis methods based on static image are not applicable. In this paper, aiming at the sequence images, with the goal of judging emotion categories and improving the accuracy of overall recognition, we study the two aspects of feature extraction and classification, focusing on the realization and improvement of feature extraction algorithm.In feature extraction, we excavate the correlation between images from two angles of texture analysis and brightness tracking. As it is know that texture generally exists in the image of this common law, the dynamic characteristics of micro-expression are extracted by using a spatio-temporal local texture description operator(LBP-TOP). Considering the differences in the degree of texture when we use the information of different directions in the image to describe it, two steps are taken to implement multi-scale texture analysis by Gaussian derivative preprocessing and LBP-TOP in different directions.Secondly, based on two constraints of illumination invariant and local smoothing, the global optical flow technique is first employed in this paper to track the change of pixel brightness between gray level images. In order to solve the problem of short duration of micro-expression, we use Gauss image pyramid and Iterative reweighted least square method to optimize the objective function between two adjacent frames, so as to estimate the optimal optical flow which can exactly reflect the motion of pixels. Due to little difference between two adjacent frames, the optical flow is too weak to reflect the subtle changes in key areas of human face, we transmit the motion information of every two frames to get the optical flow between multiple frames and form global dense optical flow field, which can rule out the low intensity of trouble. On the basis of all above, the optical flow images are divided into several regions and the features can be counted in a certain area.Further more, to offset the calculation error caused by background noise and facial skin material, thereby improving the classification effect, a novel feature combination method is proposed, which combines spatio-temporal feature extracted by LBP-TOP with optical flow feature to describe the missing details.When designing the classifier, two mature machine learning algorithms(support vector machine and random forest) are chosen to generate classification models respectively in the principle of time cost control, high efficiency and easy to use. With the purpose of distinguishing multiple categories of samples, support vector machine construct classifier in each of two categories, the classification hyper plane in which could be found by the optimal parameters through the strategy of cross validation and grid search; And random forest use two kinds of random thoughts including bagging and characteristic subspace to reduce the generalization error, the result rely on prediction voted by a combination of several decision tree models.The experimental results show that the global optical flow technique can be applied to the study of micro-expression recognition, both overall recognition accuracy and category discrimination in CASMEII database are significantly better with the improved algorithm, moreover, the method of multi-scale texture analysis has a great improvement on the recognition effect, in addition, LBP-TOP feature and optical flow feature can be complementary to each other, all these data prove that the method in this paper is reasonable and effective, and it is feasible.