FER-2013

A deep learning architecture for classifying facial emotions, learning from FER-2013 dataset.

Project Structure

fer-2013
.
├── checkpoints
├── configs
│   └── config.yaml
├── data
│   └── raw
│       ├── test
│       └── train
├── environment.yaml
├── LICENSE
├── main.py
├── notebooks
│   ├── 01_eda.ipynb
│   └── 02_evaluation.ipynb
├── README.md
├── src
│   ├── config.py
│   ├── data
│   │   ├── dataset.py
│   │   ├── fetch_data.py
│   │   └── __init__.py
│   ├── focal_loss
│   │   ├── focal_loss.py
│   │   └── __init__.py
│   └── model
│       ├── callbacks.py
│       ├── eval.py
│       ├── __init__.py
│       ├── model.py
│       └── train.py
└── tests

Setup

conda env create -f environment.yml
conda activate fer-2013

Run

python -m main

Model Performance

Custom CNN Architecture (Current)

Our custom CNN architecture achieves 65.0% accuracy on the FER-2013 test set, significantly outperforming the ResNet18 backbone variant by 11.9 percentage points.

Performance comparison:

• Custom CNN: 65.0% accuracy (F1: 0.6335)
• ResNet18 Backbone: 53.1% accuracy (F1: 0.5244)
• Improvement: +11.9% absolute accuracy gain

This demonstrates that purpose-built architectures can outperform transfer learning for specialized tasks like facial expression recognition on low-resolution (48×48) grayscale images. The custom architecture's design, optimized for FER-2013's characteristics, proves more effective than adapting a general-purpose ImageNet-pretrained backbone.

Detailed Metrics (Custom CNN)

Emotion	Precision	Recall	F1-Score	Support
Angry	0.5713	0.5939	0.5824	958
Disgust	0.6789	0.6667	0.6727	111
Fear	0.5000	0.3730	0.4273	1024
Happy	0.8692	0.8388	0.8537	1774
Neutral	0.5733	0.7105	0.6346	1233
Sad	0.5384	0.4667	0.5000	1247
Surprise	0.7034	0.8363	0.7642	831

Macro Average: Precision=0.6335, Recall=0.6409, F1=0.6335

Evaluation Results

Confusion Matrix

The confusion matrix shows the model's predictions across all emotion categories (normalized by true values):

ROC and Precision-Recall Curves

One-vs-rest ROC and precision-recall curves for each emotion class:

Training History

Loss and F1 score curves across training epochs:

Key Findings

• Best Performance: Happy emotion achieves the highest F1-score (0.8537)
• Most Challenging: Fear emotion has the lowest recall (0.3730)
• Balanced Results: The model shows consistent macro-average metrics, indicating reasonable generalization across emotion classes
• Strong Specificity: The model maintains good precision for emotions like Happy (0.8692) and Surprise (0.7034)
• Confusion between classes: classes with lowest metrics are often confounded with similar but subtetly different emotions.

Architecture Comparison: Custom CNN vs ResNet18

We experimented with replacing our custom CNN with a ResNet18 backbone to investigate whether transfer learning could improve performance. The results clearly favor the custom architecture:

Metric	Custom CNN	ResNet18	Difference
Accuracy	65.0%	53.1%	+11.9%
Weighted F1	0.6335	0.5244	+0.1091
Happy F1	0.8537	0.7686	+0.0851
Fear F1	0.4273	0.2649	+0.1624

Why Custom CNN Outperforms ResNet18

1. Resolution Mismatch: ResNet architectures are designed for higher-resolution images (224×224+), while FER-2013 uses 48×48 images. The aggressive downsampling in ResNet's early layers destroys critical facial details.

2. Task-Specific Design: Our custom architecture uses properly-sized conv blocks optimized for 48×48 input, preserving spatial information throughout the network.

3. Simpler is Better: For this constrained problem (low resolution, grayscale, limited classes), a purpose-built shallow architecture captures relevant patterns more effectively than a deep, general-purpose backbone.

4. Gabor Features: Our custom pipeline integrates Gabor filters for texture extraction, which complements the learned features better than ResNet's standard convolutions.

Conclusion: While transfer learning is powerful, domain-specific architecture design remains crucial for specialized computer vision tasks, especially with non-standard input characteristics.

License

This project is under MIT license LICENSE.