A New Feature Extraction Technique for Human Facial Expression Recognition Systems Using Depth Camera

The analysis of facial expressions in telemedicine and healthcare plays a significant role in providing sufficient information about patients like stroke and cardiac in monitoring their expressions for better management of their diseases. Due to some privacy concerns, depth camera is a good candidate in such domains over RGB camera for facial expression recognition (FER). The accuracy of such FER systems are completely reliant on the extraction of the informative features. In this work, we have tested and validated the accuracy of a new feature extraction method based on symlet wavelet transform. In this method, the human face is divided into number of regions and in each region the movement of pixels have been traced in order to create the feature vectors. Each expression frame is decomposed up to 4 levels. In each decomposition level, the distance between the two corresponding pixels is found by using the distance formula in order to extract the most informative coefficients. After feature vector creation, Linear Discriminant Analysis (LDA) has been employed to reduce the dimensions of the feature space. Lastly, Hidden Markov Model (HMM) has been exploited for expression recognition. Most of the previous FER systems used existing available standard datasets and all the datasets were pose-based datasets. Therefore, we have collected our own depth data of 15 subjects by employing the dept camera. For the whole experiments, 10-fold cross validation scheme was utilized for the experiments. The proposed technique showed a significant improvement in accuracy against the existing works.