Synchronous facial action binds dynamic facial features

Asynchrony enhances uncanniness in human, android, and virtual dynamic facial expressions

OBJECTIVE

Uncanniness plays a vital role in interactions with humans and artificial agents. Previous studies have shown that uncanniness is caused by a higher sensitivity to deviation or atypicality in specialized categories, such as faces or facial expressions, marked by configural processing. We hypothesized that asynchrony, understood as a temporal deviation in facial expression, could cause uncanniness in the facial expression. We also hypothesized that the effect of asynchrony could be disrupted through inversion.

RESULTS

Sixty-four participants rated the uncanniness of synchronous or asynchronous dynamic face emotion expressions of human, android, or computer-generated (CG) actors, presented either upright or inverted. Asynchrony vs. synchrony expressions increased uncanniness for all upright expressions except for CG angry expressions. Inverted compared with upright presentations produced less evident asynchrony effects for human angry and android happy expressions. These results suggest that asynchrony can cause dynamic expressions to appear uncanny, which is related to configural processing but different across agents.

Differences in configural processing for human versus android dynamic facial expressions

Humanlike androids can function as social agents in social situations and in experimental research. While some androids can imitate facial emotion expressions, it is unclear whether their expressions tap the same processing mechanisms utilized in human expression processing, for example configural processing. In this study, the effects of global inversion and asynchrony between facial features as configuration manipulations were compared in android and human dynamic emotion expressions. Seventy-five participants rated (1) angry and happy emotion recognition and (2) arousal and valence ratings of upright or inverted, synchronous or asynchronous, android or human agent dynamic emotion expressions. Asynchrony in dynamic expressions significantly decreased all ratings (except valence in angry expressions) in all human expressions, but did not affect android expressions. Inversion did not affect any measures regardless of agent type. These results suggest that dynamic facial expressions are processed in a synchrony-based configural manner for humans, but not for androids.

Modulation transfer functions for audiovisual speech

Temporal synchrony between facial motion and acoustic modulations is a hallmark feature of audiovisual speech. The moving face and mouth during natural speech is known to be correlated with low-frequency acoustic envelope fluctuations (below 10 Hz), but the precise rates at which envelope information is synchronized with motion in different parts of the face are less clear. Here, we used regularized canonical correlation analysis (rCCA) to learn speech envelope filters whose outputs correlate with motion in different parts of the speakers face. We leveraged recent advances in video-based 3D facial landmark estimation allowing us to examine statistical envelope-face correlations across a large number of speakers (∼4000). Specifically, rCCA was used to learn modulation transfer functions (MTFs) for the speech envelope that significantly predict correlation with facial motion across different speakers. The AV analysis revealed bandpass speech envelope filters at distinct temporal scales. A first set of MTFs showed peaks around 3-4 Hz and were correlated with mouth movements. A second set of MTFs captured envelope fluctuations in the 1-2 Hz range correlated with more global face and head motion. These two distinctive timescales emerged only as a property of natural AV speech statistics across many speakers. A similar analysis of fewer speakers performing a controlled speech task highlighted only the well-known temporal modulations around 4 Hz correlated with orofacial motion. The different bandpass ranges of AV correlation align notably with the average rates at which syllables (3-4 Hz) and phrases (1-2 Hz) are produced in natural speech. Whereas periodicities at the syllable rate are evident in the envelope spectrum of the speech signal itself, slower 1-2 Hz regularities thus only become prominent when considering crossmodal signal statistics. This may indicate a motor origin of temporal regularities at the timescales of syllables and phrases in natural speech.

Automated Facial Expression Recognition Framework Using Deep Learning

Facial expression is one of the most significant elements which can tell us about the mental state of any person. A human can convey approximately 55% of information nonverbally and the remaining almost 45% through verbal communication. Automatic facial expression recognition is presently one of the most difficult tasks in the computer science field. Applications of facial expression recognition (FER) are not just limited to understanding human behavior and monitoring person's mood and the mental state of humans. It is also penetrating into other fields such as criminology, holographic, smart healthcare systems, security systems, education, robotics, entertainment, and stress detection. Currently, facial expressions are playing an important role in medical sciences, particularly helping the patients with bipolar disease, whose mood changes very frequently. In this study, an algorithm, automated framework for facial detection using a convolutional neural network (FD-CNN) is proposed with four convolution layers and two hidden layers to improve accuracy. An extended Cohn-Kanade (CK+) dataset is used that includes facial images of different males and females with expressions such as anger, fear, disgust, contempt, neutral, happy, sad, and surprise. In this study, FD-CNN is performed in three major steps that include preprocessing, feature extraction, and classification. By using this proposed method, an accuracy of 94% is obtained in FER. In order to validate the proposed algorithm, K-fold cross-validation is performed. After validation, sensitivity and specificity are calculated which are 94.02% and 99.14%, respectively. Furthermore, the f1 score, recall, and precision are calculated to validate the quality of the model which is 84.07%, 78.22%, and 94.09%, respectively.