Recognition of facial action units from video streams with recurrent neural networks : a new paradigm for facial expression recognition
This research investigated the application of recurrent neural networks (RNNs) for recognition of facial expressions based on facial action coding system (FACS). Support vector machines (SVMs) were used to validate the results obtained by RNNs. In this approach, instead of recognizing whole facial expressions, the focus was on the recognition of action units (AUs) that are defined in FACS. Recurrent neural networks are capable of gaining knowledge from temporal data while SVMs, which are time invariant, are known to be very good classifiers. Thus, the research consists of four important components: comparison of the use of image sequences against single static images, benchmarking feature selection and network optimization approaches, study of inter-AU correlations by implementing multiple output RNNs, and study of difference images as an approach for performance improvement. In the comparative studies, image sequences were classified using a combination of Gabor filters and RNNs, while single static images were classified using Gabor filters and SVMs. Sets of 11 FACS AUs were classified by both approaches, where a single RNN/SVM classifier was used for classifying each AU. Results indicated that classifying FACS AUs using image sequences yielded better results than using static images. The average recognition rate (RR) and false alarm rate (FAR) using image sequences was 82.75% and 7.61%, respectively, while the classification using single static images yielded a RR and FAR of 79.47% and 9.22%, respectively. The better performance by the use of image sequences can be at- tributed to RNNs ability, as stated above, to extract knowledge from time-series data. Subsequent research then investigated benchmarking dimensionality reduction, feature selection and network optimization techniques, in order to improve the performance provided by the use of image sequences. Results showed that an optimized network, using weight decay, gave best RR and FAR of 85.38% and 6.24%, respectively. The next study was of the inter-AU correlations existing in the Cohn-Kanade database and their effect on classification models. To accomplish this, a model was developed for the classification of a set of AUs by a single multiple output RNN. Results indicated that high inter-AU correlations do in fact aid classification models to gain more knowledge and, thus, perform better. However, this was limited to AUs that start and reach apex at almost the same time. This suggests the need for availability of a larger database of AUs, which could provide both individual and AU combinations for further investigation. The final part of this research investigated use of difference images to track the motion of image pixels. Difference images provide both noise and feature reduction, an aspect that was studied. Results showed that the use of difference image sequences provided the best results, with RR and FAR of 87.95% and 3.45%, respectively, which is shown to be significant when compared to use of normal image sequences classified using RNNs. In conclusion, the research demonstrates that use of RNNs for classification of image sequences is a new and improved paradigm for facial expression recognition.