ademaulana/vitDeteksiEkspresi

Vision Transformer (ViT) for Facial Expression Recognition Model Card

Model Overview

• Model Name: trpakov/vit-face-expression
• Task: Facial Expression/Emotion Recognition
• Dataset: FER-2013
• Model Architecture: Vision Transformer (ViT)
• Finetuned from model: vit-base-patch16-224-in21k

Model Description

The vit-face-expression model is a Vision Transformer fine-tuned for facial emotion recognition using the FER2013 dataset. The dataset consists of facial images categorized into seven emotions:

• Angry
• Disgust
• Fear
• Happy
• Sad
• Surprise
• Neutral

This model leverages transformer-based architecture to capture spatial dependencies in facial features effectively.

Data Preprocessing

Before feeding images into the model, the following preprocessing steps are applied:

• Resizing: Images are resized to 224x224 pixels.
• Normalization: Pixel values are normalized using the ViT feature extractor's predefined mean and standard deviation.
• Data Augmentation: The following transformations are applied to improve generalization:
  • Random horizontal flips (p=0.5)
  • Random rotation (±10 degrees)
  • Color jitter (brightness and contrast adjustments)

Training Details

• Optimizer: AdamW
• Learning Rate: 5e-5
• Batch Size: 8
• Loss Function: Cross-Entropy Loss
• Epochs: 3
• Imbalanced Data Handling: The dataset was limited to a maximum of 5000 samples per class to maintain balance.

Evaluation Metrics

Metric	Value
Validation Accuracy	0.7113
Test Accuracy	0.7116

Confusion Matrix Insights

The model exhibits the following patterns:

• Neutral expressions are frequently misclassified, often leading to incorrect predictions.
• Disgust and Fear have lower recall rates, meaning the model struggles to distinguish these emotions accurately.
• Happy is often confused with Neutral and Fear, likely due to overlapping facial features.
• Angry and Sad also show some misclassification overlap, which might be addressed with better feature extraction.

Limitations

• Data Bias: The model's performance is influenced by the class imbalance in the FER2013 dataset.
• Generalization Issues: The model may not generalize well to diverse facial expressions due to dataset limitations.
• Misclassification Trends: Higher misclassification rates for minority classes such as 'Disgust' and 'Fear' suggest potential areas for improvement.