An Attention-Enhanced ConvNeXtTiny Model for Robust Facial Expression Recognition

Analgesia; Facial Expression Recognition, ConvNeXtTiny, CBAM, Deep Learning, Transfer Learning, Attention Mechanism, FER2013, RAF-DB.

1*Ms. Manisha B. Thombare
Research Scholar, Department of Computer
Engineering,MET’s BKC, Institute of
Engineering, Nashik, Maharashtra, India
Savitribai Phule Pune University, Pune,
Maharashtra, India
manishathombare27@gmail.com

2Dr. Shyamrao V. Gumaste
Professor and Head, Department of Artificial
Intelligence and Data Science,
MET’s BKC, Institute of Engineering,
Nashik, Maharashtra, India
Savitribai Phule Pune University, Pune,
Maharashtra, India, Email:
svgumaste@gmail.com

Received:26-06-2026
Revised:29-07-2026
Accepted:04-08-2026

European Journal of Prosthodontics and Restorative Dentistry (2026) 34 (7s), 870–887

An Attention-Enhanced ConvNeXtTiny Model for Robust Facial Expression Recognition

Facial Expression Recognition (FER) plays an important role in affective computing and human–computer interaction by enabling automated interpretation of human emotional states from facial images. Despite recent advances in deep learning, reliable FER remains challenging because of variations in facial appearance, illumination, pose, occlusion, image quality, class imbalance, and visual similarity between emotion categories. This study proposes a facial expression recognition model that combines the ConvNeXtTiny architecture with the Convolutional Block Attention Module (CBAM) to improve discriminative facial feature learning. ConvNeXtTiny is employed to extract hierarchical facial representations, while CBAM refines these representations through complementary channel and spatial attention. Transfer learning, data augmentation, label smoothing, and adaptive optimization are incorporated during model training to improve convergence and generalization. The proposed model is evaluated on the FER2013 and RAF-DB benchmark datasets using accuracy, precision, recall, F1-score, confusion matrix analysis, cross-validation, computational complexity, and inference-time analysis. The experimental results demonstrate competitive recognition performance across both datasets. On FER2013, the model achieves an accuracy of 77.0%, precision of 77.5%, recall of 77.0%, and F1-score of 77.2%. On RAF-DB, the model achieves an accuracy of approximately 92.9%, demonstrating its capability to recognize facial expressions under diverse real-world conditions. Ablation experiments further indicate the contribution of the attention mechanism and optimization components to the final recognition performance. The findings demonstrate that combining modern convolutional feature extraction with lightweight channel-spatial attention provides an effective approach for facial expression recognition.

•••••••••••••••••••••••••••••••• ejprd.org - Published by Riset Publication Services LLC

EJPRD

Copyright ©2026 by Riset Publication Services LLC