Facial expression recognition sensing the complexity of testing samples

Facial expression recognition has always been a challenging issue due to the inconsistencies in the complexity of samples and variability of between expression categories. Many facial expression recognition methods train a classification model and then use this model to identify all test samples, without considering the complexity of each test sample. They are inconsistent with human cognition laws such as the principle of simplicity, so that they are easily under-learned and then are difficult to identify test samples correctly. Hence, this paper proposed a new facial expression recognition method sensing the complexity of test samples, which can nicely solve the problem of the inconsistent distribution of samples complexity. It firstly divided the training data into the hard subset and the easy subset for classification according to the complexity of samples for expression recognition. Subsequently, these two subsets are applied to train two classifiers. Instead of using the same classifier to predict all test samples, our method assigned each test sample to the corresponding classifier based on the complexity of the test sample. The experimental results demonstrated the effectiveness of the proposed method and obtained a significant improvements of the recognition performance on benchmark datasets.