Analyze Facial Expression Recognition Based on Curvelet Transform via Extreme Learning Machine

摘要

This paper aims to investigate the key factors of facial expression recognition based on local curvelet transform for real-time training data. Local curvelet transform (LCT) is the application of curvelet transform that benefits from useful features extracted by curvelet transform and reduces the computation cost of using all curvelet coefficients. The reduction of computation is through calculating the representative features, instead of directly using all curvelet coefficients. The representative features are mean, standard deviation and entropy. This approach has been reported to achieve impressively 0.9445 and 0.9486 accuracy on JAFFE and Cohn-Kanade datasets. However, there are many factors influencing the final performance, in which these factors have not been thoroughly studied. Our investigation has shown that these factors could result up to almost 10% difference and their effects are thoroughly studied.