A Slim Convolutional Neural Network for Low-cost Masked Facial Expression Classifier

The facial expression emphasizes a predictive facial gesture widely used for emotion recognition. The pandemic situation encourages the demand for the use of masks in specific areas, urging a challenge from a facial analysis method to accurately predict limited facial features. Practical applications demand a classification system that operates quickly. This paper proposes an accurate real-time facial expression recognition for masked faces. It offers an efficient backbone that improves from MobileNetV2, which drives the model to generate parameters and less computation. The represented context module is introduced to capture the attention of interest information and highlight features globally without placing a significant computational burden. In order to emphasize a reliable classification system in real applications, it proposes the development of a masked facial expression dataset, namely M-KDEF. The experimental results show that the facial expression classifier performs compared to other methods with an accuracy of 96.99% on the M-KDEF and 87.14% on the M-LFW dataset. It also examines the speed of the integrated system that obtains a data processing speed of 61.30 frames per second on the CPU.