Facial expression recognition (FER) is crucial in expressing a human's emotional state. Emotions and expressions on the human face are information that computers and deep learning can recognize. FER is a current research topic due to the recent advancement and use of human-computer interaction systems. The recognition of facial expressions is challenging for current deep learning models due to the variable brightness, background, pose, etc. of the face images. This work presents an improved learning method based on a feature fusion convolutional neural network (CNN) to recognize seven facial expressions. First, we trained three different base model structures and then fused the features of the pre-trained base models 1 and 2 from the final fully connected layers to obtain a fusion network. Second, the fusion network was trained and used with the third pre-trained base model for performance evaluation. We applied the max-score fusion and mean-score fusion techniques between the fusion network and the third pre-trained base model to predict the output class. Our results indicated that the proposed method outperforms the base models in all metrics and achieves a classification accuracy of 69.03% on the facial expression dataset.
Ahmed et al. (Mon,) studied this question.