Facial Emotion Recognition Using ResNet50
Abstract
This work focuses on constructing a deep learning-based model for facial emotion recognition, taking ResNet50 as the backbone. To boost generalization, a number of data augmentation techniques, including rotating, zooming, shifting, and flipping, have significantly increased the dataset of 152 images that represent eight emotion classes to 7,752 samples. The labels were then stored in a categorical format, and the images were scaled to 224 × 224 pixels. The dataset was divided into two sections: a test set with 1,551 samples and a training set with 6,201 samples. In order to prevent overfitting, it was decided to learn the model using a batch size of 32, early halting, and a method for reducing the learning rate. It was actually very good, with a loss value of 0.197 and a test accuracy of 94.8%. A classification report was performed, which verified that the performance was well-balanced among all classes. Also, the confusion matrix validated the good recognition with minor misclassifications. This framework, hence, shows a robust and efficient emotion recognition capability.