Improve the uneven distribution of micro-expression recognition data

Abstract

Micro-expressions are the most intuitive response to human emotions. In this research, we will deeply study the recognition of micro-expressions and classify each class of expressions, and use an end-to-end deep learning method to solve the problem of data uneven caused by the uneven number of different types of expressions in micro-expression recognition. The recognition model VGG16 based on the deep convolutional neural network is used to realize micro-expression recognition. Because the deep convolutional neural network can automatically extract the characteristics of hierarchical features, which makes it more and more widely used in many machine learning applications. A large amount of data is required to train a deep convolutional neural network, but due to the difficulty in collecting micro-expression image, result the training data for the micro-expression recognition model is insufficient. The model parameters that have been trained by VGG16 are used to introduce a new model to help the training of the new model, and the Focal Loss function is introduced as the objective function to reduce the impact of data uneven. Thereby, the accuracy of micro-expressions recognition can be improved and achieve the improvement of the classification performance of micro-expression.