DEEP LEARNING BASED APPROACH FOR FACIAL EXPRESSION RECOGNITION

Authors

  • Mahnoor Zaheer (mahnoorzaheer001@gmail.com) Department of Computer Science and Information Technology, The Superior University, Lahore 54000, Pakistan Author
  • H. M. Shahzad( shahzad.dar@superior.edu.pk) Author

Keywords:

facial expression recognition; EfficientNet-B3; CBAM; deep learning; transfer learning; FER2013; RAF-DB; attention mechanism; class imbalance; annotation quality

DOI:

1102/3525

Abstract

Recognizing emotion from facial images sits at the center of affective computing, with downstream value for human computer interaction, mental  health monitoring, and behavioral surveillance.  This study proposes a deep learning method that uses EfficientNet-B3 to extract features from images, refined by a Convolutional Block Attention Module (CBAM) that attends to both channel and spatial detail. Images are prepared through a three-step pipeline edge preserving denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE), and adaptive gamma correction — to improve image quality before feature extraction. The proposed framework is evaluated on two publicly available, officially sourced benchmarks that share the same real world characteristic of severe class imbalance: the FER2013 dataset and the unbalanced RAF-DB dataset, obtained directly from the dataset authors. Using the same architecture, training method, and class-imbalance management on both datasets, the suggested model obtains test accuracies of 68.54% on FER2013 and 83.15% on RAF-DB, a difference of 14.61 percentage points. Because both benchmarks retain their natural, unaltered class distribution, this gap is attributable primarily to differences in annotation quality and image-collection protocol rather than to class balance, isolating a factor that is frequently confounded in the FER literature. Rather than proposing a new architecture, this work's contribution is empirical: it isolates the effect of dataset annotation quality from model capability by holding the architecture, training procedure, and imbalance-handling strategy constant across two benchmarks of differing annotation reliability.

Downloads

Published

2026-03-03