Emotion Recognition From Facial Expression Using Deep Learning

Abstract

Facial emotion recognition (FER) is one of the major parts of affective computing, which enables machines to read the affective states of human faces in photographed faces[1][2]. Amazing recent advances in such vision tasks have been made by deep convolutional neural networks (CNNs)[1][3]. The paper is a contribution in which we create and test a CNN-based model to perform the classification of seven fundamental emotions (anger, disgust, fear, happiness, sadness, surprise, neutral) using face images. Our model uses a sequence of convolutional and pooling layers (see Figure 1) to automatically learn discriminative features from preprocessed face images, followed by fully-connected layers for final classification. We augment the FER2013 benchmark dataset (≈35K 48×48 grayscale images) with data augmentation and careful preprocessing to address class imbalance and noise[4][5]. The proposed CNN achieves ~67% accuracy on FER2013, which is comparable to state-of-the-art reported results (e.g. 65–73% range[6]) and confirms the effectiveness of deep feature learning. We also analyze model interpretability via class activation mapping. The main contributions of this work are: (1) a robust CNN architecture tailored for FER in the wild, incorporating preprocessing and augmentation strategies to improve generalization[4][7]; (2) extensive evaluation on FER2013, demonstrating competitive accuracy; and (3) insight into feature representations and error patterns through visualization of learned filters and confusion matrices. These results advance the automation of facial affect analysis in practical applications.