Hybrid MobileNetV3-LSTM model for detecting phasic and tonic receptor responses in facial images

Abstract

Pain was a complex and subjective experience that involved sensory, emotional, and cognitive aspects simultaneously. The responses of phasic and tonic receptors to painful stimuli exhibited different patterns and could be observed through facial expressions as a form of nonverbal communication. This study aimed to implement a MobileNetV3-LSTM model to classify phasic, tonic, and normal receptor responses using human facial expression images. The objective was to obtain the most optimal model for classifying facial expressions exposed to pain stimuli targeting phasic or tonic receptors. The methods involved the development and evaluation of three models: MobileNetV3Large, MobileNetV3Small, and their respective hybrid versions combined with LSTM, to examine the effect of incorporating temporal information on classification performance. According to results 10, the hybrid MobileNetV3Large-LSTM model performed the best, with an F1-score of 94%, accuracy of 93%, precision of 96%, and recall of 93% on the 12 test data. Meanwhile, the MobileNetV3Small-LSTM model reached 74% accuracy, 80% precision, 74% recall, and a 74% F1-score. The standalone MobileNetV3Large model only obtained 68% accuracy and an F1-score of 0.59, while MobileNetV3Small without LSTM achieved 75% accuracy and an F1-score of 0.56. These results suggest that the inclusion of LSTM layers greatly enhanced the accuracy in the model. This research added to the development of facial expression classification methodologies to recognize pain and complemented the body of knowledge on hybrid model utilization in deep learning.