Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Facial Emotion Recognition on FER-2013 using an EfficientNetB2-Based Approach

Sahil Naik, Soham Bagayatkar, Pavankumar Singh

TL;DR

The paper tackles Facial Emotion Recognition on the FER-2013 dataset, addressing noise and imbalance with a compact EfficientNetB2-based pipeline designed for real-time use. A two-stage warmup and fine-tuning regime, combined with AdamW optimization, label smoothing, and clipped class weights, stabilizes learning in the presence of noisy labels and rare classes. On the official test set, it achieves 68.78% accuracy with about 9.2M parameters, demonstrating a strong accuracy/efficiency trade-off compared to large VGG-like models. The authors provide a reproducible framework with detailed per-class analyses, confusion matrices, and ablation guidance, underscoring the approach's practicality for edge and real-time FER applications. Overall, the work advances Green AI in FER by showing how careful design and optimization of a compact architecture can rival larger models while enabling deployment on resource-limited devices.

Abstract

Detection of human emotions based on facial images in real-world scenarios is a difficult task due to low image quality, variations in lighting, pose changes, background distractions, small inter-class variations, noisy crowd-sourced labels, and severe class imbalance, as observed in the FER-2013 dataset of 48x48 grayscale images. Although recent approaches using large CNNs such as VGG and ResNet achieve reasonable accuracy, they are computationally expensive and memory-intensive, limiting their practicality for real-time applications. We address these challenges using a lightweight and efficient facial emotion recognition pipeline based on EfficientNetB2, trained using a two-stage warm-up and fine-tuning strategy. The model is enhanced with AdamW optimization, decoupled weight decay, label smoothing (epsilon = 0.06) to reduce annotation noise, and clipped class weights to mitigate class imbalance, along with dropout, mixed-precision training, and extensive real-time data augmentation. The model is trained using a stratified 87.5%/12.5% train-validation split while keeping the official test set intact, achieving a test accuracy of 68.78% with nearly ten times fewer parameters than VGG16-based baselines. Experimental results, including per-class metrics and learning dynamics, demonstrate stable training and strong generalization, making the proposed approach suitable for real-time and edge-based applications.

Facial Emotion Recognition on FER-2013 using an EfficientNetB2-Based Approach

TL;DR

The paper tackles Facial Emotion Recognition on the FER-2013 dataset, addressing noise and imbalance with a compact EfficientNetB2-based pipeline designed for real-time use. A two-stage warmup and fine-tuning regime, combined with AdamW optimization, label smoothing, and clipped class weights, stabilizes learning in the presence of noisy labels and rare classes. On the official test set, it achieves 68.78% accuracy with about 9.2M parameters, demonstrating a strong accuracy/efficiency trade-off compared to large VGG-like models. The authors provide a reproducible framework with detailed per-class analyses, confusion matrices, and ablation guidance, underscoring the approach's practicality for edge and real-time FER applications. Overall, the work advances Green AI in FER by showing how careful design and optimization of a compact architecture can rival larger models while enabling deployment on resource-limited devices.

Abstract

Detection of human emotions based on facial images in real-world scenarios is a difficult task due to low image quality, variations in lighting, pose changes, background distractions, small inter-class variations, noisy crowd-sourced labels, and severe class imbalance, as observed in the FER-2013 dataset of 48x48 grayscale images. Although recent approaches using large CNNs such as VGG and ResNet achieve reasonable accuracy, they are computationally expensive and memory-intensive, limiting their practicality for real-time applications. We address these challenges using a lightweight and efficient facial emotion recognition pipeline based on EfficientNetB2, trained using a two-stage warm-up and fine-tuning strategy. The model is enhanced with AdamW optimization, decoupled weight decay, label smoothing (epsilon = 0.06) to reduce annotation noise, and clipped class weights to mitigate class imbalance, along with dropout, mixed-precision training, and extensive real-time data augmentation. The model is trained using a stratified 87.5%/12.5% train-validation split while keeping the official test set intact, achieving a test accuracy of 68.78% with nearly ten times fewer parameters than VGG16-based baselines. Experimental results, including per-class metrics and learning dynamics, demonstrate stable training and strong generalization, making the proposed approach suitable for real-time and edge-based applications.
Paper Structure (25 sections, 4 figures, 3 tables)

This paper contains 25 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Positioning vs. State-of-the-Art
  4. Dataset and Preprocessing
  5. FER-2013 Overview
  6. Train/Validation/Test Split
  7. Preprocessing and Augmentation
  8. Methodology
  9. Model Architecture
  10. Training Regime
  11. Regularization and Callbacks
  12. Experimental Setup
  13. Results
  14. Learning Dynamics
  15. Test Performance
  16. ...and 10 more sections

Figures (4)

  • Figure 1: It is a depiction of Training vs. validation accuracy compared with an epoch. The close alignment of curves indicates stable learning and good generalization.
  • Figure 2: Training vs. validation loss across epochs, showing smooth convergence and stable optimization.
  • Figure 3: Confusion matrix (absolute counts) on the FER-2013 test set.
  • Figure 4: Row-normalized confusion matrix showing per-class recall.