Facial Emotion Recognition of 16 Distinct Emotions From Smartphone Videos: Comparative Study of Machine Learning and Human Performance.

Background: The development of automatic emotion recognition models from smartphone videos is a crucial step toward the dissemination of psychotherapeutic app interventions that encourage emotional expressions. Existing models focus mainly on the 6 basic emotions while neglecting other therapeutically relevant emotions. To support this research, we introduce the novel Stress Reduction Training Through the Recognition of Emotions Wizard-of-Oz (STREs WoZ) dataset, which contains facial videos of 16 distinct, therapeutically relevant emotions.

Objective: This study aimed to develop deep learning-based automatic facial emotion recognition (FER) models for binary (positive vs negative) and multiclass emotion classification tasks, assess the models' performance, and validate them by comparing the models with human observers.

Methods: The STREs WoZ dataset contains 14,412 facial videos of 63 individuals displaying the 16 emotions. The selfie-style videos were recorded during a stress reduction training using front-facing smartphone cameras in a nonconstrained laboratory setting. Automatic FER models using both appearance and deep-learned features for binary and multiclass emotion classification were trained on the STREs WoZ dataset. The appearance features were based on the Facial Action Coding System and extracted with OpenFace. The deep-learned features were obtained through a ResNet50 model. For our deep learning models, we used the appearance features, the deep-learned features, and their concatenation as inputs. We used 3 recurrent neural network (RNN)-based architectures: RNN-convolution, RNN-attention, and RNN-average networks. For validation, 3 human observers were also trained in binary and multiclass emotion recognition. A test set of 3018 facial emotion videos of the 16 emotions was completed by both the automatic FER model and human observers. The performance was assessed with unweighted average recall (UAR) and accuracy.

Results: Models using appearance features outperformed those using deep-learned features, as well as models combining both feature types in both tasks, with the attention network using appearance features emerging as the best-performing model. The attention network achieved a UAR of 92.9% in the binary classification task, and accuracy values ranged from 59.0% to 90.0% in the multiclass classification task. Human performance was comparable to that of the automatic FER model in the binary classification task, with a UAR of 91.0%, and superior in the multiclass classification task, with accuracy values ranging from 87.4% to 99.8%.

Conclusions: Future studies are needed to enhance the performance of automatic FER models for practical use in psychotherapeutic apps. Nevertheless, this study represents an important first step toward advancing emotion-focused psychotherapeutic interventions via smartphone apps.