Comparative Analysis of Pre-Trained Models and Interpolation for Facial Expression Recognition

Abstract

This study presents a comprehensive examination of facial emotion detection using the FER 2013 dataset. In numerous disciplines, including psychology, human-computer interface, and marketing research, efficient facial expression recognition is a crucial problem to resolve. This study experimented with four different convolutional neural network architectures: ResNet50, VGG16, EfficientNetV2L, and NasNetLarge, in order to categorize seven different emotions: anger, disgust, fear, happiness, sadness, surprise, and neutral. Before passing the images to the classification layers, this study feeds the images to ResNet50, VGG16, EfficientNetV2L, and NasNetLarge to extract features from them. Model and interpolation methods are validated and assessed using the Public Text in two different sizes (128,128) and (256,256). The Adam optimizer is used to train models, with a batch size of 32 and a learning rate 0.001. The accuracy of the models is assessed during the evaluation process. In order to determine whether interpolation produces better results or not, this study will conduct an experimental analysis of the interpolation used to upscale a lower-resolution image. Additionally, the most recent state-of-the-art models need images with higher resolution, so using interpolation to upscale images could enable the use of models like NasNetLarge and others. For all considered models this study could get at most 69% validation accuracy using interpolation to resize the images from 48x48 to 128x128 and 256x256. This study provides insightful viewpoints of the effectiveness of different pre-trained and publicly accessible CNN architectures for feature extraction and interpolation methods in facial emotion detection. The findings can direct the selection of appropriate models with appropriate interpolation sizes for emotion detection applications and stimulate additional study in this area.