Facial Expression Recognition using deep Convolutional Neural Network

Abstract

The system that reads facial emotions from images is called Facial Expression Recognition (FER) or Facial Recognition (FR). This task provides insight into an individual's emotional condition. Deep learning-based FER is a field of AI approaches that is developing quickly. the goal of facial expression analysis is to recognize human emotions. This method automatically recognizes and extracts discriminative elements from face images using deep neural networks, namely CNNs. FER systems are now far more accurate and resilient thanks to deep learning models. Large-scale annotated datasets are used to train these models. Seven pre-trained deep learning models—InceptionResNetV2, InceptionV3, MobileNetV2, ResNet101, ResNet50, Xception, and GoogleNet—were used for the FER system in this thesis. Each utilized model varies based on modified parameters. We used both the Adam and SGD optimizers to evaluate each model based on a range of batch sizes, including 16, 32, 64, and 128. The FER-2013 large dataset, which consists of small-sized grayscale images, was used to assess the models. For the purpose of feeding the deeply pre-trained deep learning models, we transform the dataset to RGB color space. We obtained 65.47%, 65.89%, 64.25%, and 65.05% for metrics accuracy, precision, recall, and f1-score, respectively, in the InceptionResNetV2 model using the Adam optimizer and batch size 128. When compared to the most advanced model, we may improve the model's correctness in terms of performance.