Original scientific paper
https://doi.org/10.32985/ijeces.17.9.2
Facial Expression Recognition Using CNNs and Attention Mechanisms on Modified FER2013 Dataset
Aqeel Majeed Breesam
; Institute of Medical Technology/Baghdad Middle Technical University Baghdad, Iraq
*
Roaa Safi Abed Alah
; Department of Intelligent Medical Systems University of Information Technology and Communications Baghdad, Iraq
Mina Faris Ali Alnaimy
; College of Engineering, Department of Media and Communications Technology University of Information Technology and Communications Baghdad, Iraq
* Corresponding author.
Abstract
Facial Expression Recognition (FER) is an important component of affective computing and human-computer interaction, yet it is difficult to apply in practice because of the variations in illumination, pose variation, occlusions, low-resolution images, and skew in the classes. To overcome these issues, this paper introduces FER-CBAMNet, a small and attention-based convolutional neural network that uses a lightweight residual network with Convolutional Block Attention Modules (CBAM) to focus on discriminative facial regions, including the eyes, eyebrows, and mouth, and ignore irrelevant background information. The model is trained with a powerful pipeline that includes a large amount of data augmentation, normalization, and class-balanced regularization on the FER-2013 dataset under a modified evaluation protocol. Experimental evidence shows that FER-CBAMNet has a mean accuracy of 96.58% and a macro F1-score of 0.963, with low inter-run variability, significantly surpassing current state-of-the-art models on the same test set, such as standard CNNs, lightweight CNN-KNN hybrids (~75%), and transformer-based models (~71\). Ablation experiments reveal that the addition of CBAM adds significant performance improvement, enhances accuracy, and balances the recognition of all seven emotion categories. Also, Grad-CAM visualizations demonstrate that the model makes decisions based on semantically meaningful parts of the face, which are interpretable and diagnostically meaningful. These results suggest that attention-enhanced lightweight CNNs are highly accurate, robust, and transparent, and FER-CBAMNet is highly suitable for real-world and resource-limited applications.
Keywords
facial expression recognition; convolutional neural network; deep learning; FER2013;
Hrčak ID:
352137
URI
Publication date:
14.10.2026.
Visits: 0 *