Facial Expression Recognition Using Improvement of ResNet-18

Facial expression recognition is a current trend in computer vision to identify human emotional states from facial cues. There are six basic human facial expressions: anger, disgust, fear, happiness, surprise, and sadness. Popular architectures like ResNet need development to improve accuracy and reduce parameter utilization. This study proposes a CNN architecture employing the ResNet model to recognize features of human facial expressions. It should be improved to reduce computation and an excessive number of parameters that could affect the accuracy and effectiveness of feature extraction. A Depth-Wise Convolution could be embedded within four stages of convolution inside ResNet to achieve the desired result. This model consists of several stages, including filtering, normalization, activation, and four stages of convolution with Depthwise embedded inside, and culminating it in a final layer with softmax activation. The architecture is trained using the CK+48 dataset to achieve accurate classification. An improved model could reach 100% validation accuracy and 0.0062 validation error, generating 0,1996 GFLOPS .