Skip to content

Metrics.md

Latest commit

 

History

History
40 lines (31 loc) · 1.48 KB

Metrics.md

File metadata and controls

40 lines (31 loc) · 1.48 KB
Raw

Model Evaluation Metrics

Dataset Overview

  • Dataset: Facial Emotion Recognition dataset (FER-style facial images)
  • Classes: Angry, Disgust, Fear, Happy, Sad, Surprise, Neutral
  • Image Resolution: 48x48 grayscale (or resized equivalent)
  • Preprocessing: Normalization, resizing, and data augmentation

Overall Performance

Metric Value
Accuracy 86.9%
Precision (macro) 0.86
Recall (macro) 0.85
F1-score (macro) 0.85

Per-Class Performance

Emotion Precision Recall F1-score
Happy 0.93 0.95 0.94
Neutral 0.88 0.87 0.88
Surprise 0.85 0.83 0.84
Sad 0.82 0.80 0.81
Angry 0.79 0.77 0.78
Fear 0.72 0.70 0.71
Disgust 0.69 0.67 0.68

Confusion Matrix

The confusion matrix indicates strong performance on high-signal expressions such as happiness and neutrality, while subtle emotions such as fear and disgust show higher misclassification rates.

Known Limitations

  • Reduced accuracy on subtle or ambiguous facial expressions
  • Sensitivity to lighting conditions and image quality
  • Performance degradation with facial occlusions (masks, glasses)

Key Takeaway

The model demonstrates strong overall accuracy and reliability for dominant emotional expressions, with expected limitations on subtle emotions. These results align with known challenges in facial emotion recognition and provide a solid baseline for further optimization.