Skip to content

MODEL_CARD.md

Latest commit

 

History

History
210 lines (141 loc) · 4.32 KB

MODEL_CARD.md

File metadata and controls

210 lines (141 loc) · 4.32 KB

🧠 MODEL_CARD.md — Facial Emotion Recognition System

Author: Corey Leath (Trojan3877)

Version: v1.0

Last Updated: 2025-11-15

📌 1. Model Overview

The Facial Emotion Recognition (FER) Model is a Convolutional Neural Network trained on the FER-2013 dataset to classify human facial expressions into seven emotion categories:

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

This model is deployed using FastAPI, Streamlit, Docker, and Kubernetes, and tracked using MLflow. Prediction logs are stored using Snowflake for enterprise-grade MLOps.

🎯 2. Intended Use

✔ Primary use case

To provide real-time emotion recognition for:

  • Education tools
  • Interactive AI demos
  • User behavior research
  • Human–computer interaction (HCI)
  • Wellness monitoring applications
  • Social robotics
  • UX or product analytics

⚠ Not intended for:

  • Law enforcement
  • Hiring decision tools
  • Predicting personality
  • Medical diagnosis
  • Legal or governmental use
  • Automated surveillance
  • High-stakes decision-making

This model should never be used as the sole decision-maker for any sensitive task.

🧩 3. Model Architecture

  • Architecture: CNN (Convolutional Neural Network)
  • Framework: TensorFlow/Keras
  • Input: 48×48 grayscale images
  • Output: 7-class softmax probability distribution

Logged in MLflow:

✔ Model architecture ✔ Parameters ✔ Metrics ✔ Training loss curves ✔ Confusion matrix ✔ Artifacts (history.json)

📚 4. Training Data

Dataset: FER-2013 (public Kaggle dataset)

Pros:

  • Large dataset (~35,000 images)
  • Multiple classes
  • Standard benchmark in Computer Vision

Cons:

  • Dataset skews toward younger faces
  • Overrepresentation of certain ethnic groups
  • Limited lighting variations
  • Some images mislabeled
  • Grayscale only

⚠️ 5. Ethical Considerations & Known Biases

Emotion recognition is an area with well-documented ethical risks, including:

1. Demographic Bias

Studies show FER datasets often underrepresent:

  • Darker skin tones
  • Older adults
  • Certain ethnic backgrounds

This can cause misclassification.

2. Cultural Expression Variability

Emotions are expressed differently across:

  • Culture
  • Gender
  • Region
  • Social norms

This model does not account for cultural variation.

3. Misuse / Overuse

FER models can be misused in:

  • Workplace surveillance
  • Student behavior scoring
  • Criminal profiling
  • Automated “emotion-based decisions”

This model must not be deployed in these contexts.

📊 6. Performance & Evaluation

Metrics Logged in MLflow

  • Accuracy
  • Validation accuracy
  • Precision, Recall, F1-score (optional)
  • Confusion matrix
  • Loss curves

Typical Performance (FER-2013 Benchmarks)

(Note: These are reference-level until you run eval)

  • Overall Accuracy: 0.64–0.72 typical range
  • Best accuracy tends to be on: Happy, Surprise
  • Lowest accuracy typically on: Disgust, Fear, Angry

You may include exact numbers after running your model evaluation script.

🧪 7. Limitations

This model struggles with:

  • Occluded faces (masks, glasses, hats)
  • Low lighting
  • Extreme angles
  • Blurry images
  • Cultural expression differences
  • Mixed or ambiguous expressions
  • Micro-expressions
  • Faces not centered or cropped

The model expects clean, centered, front-facing 48×48 grayscale faces.

🧱 8. Deployment Details

Infrastructure

  • FastAPI for inference
  • Streamlit for UI
  • Docker containerization
  • Kubernetes deployment with LoadBalancer
  • MLflow experiment tracking
  • Snowflake prediction logging
  • GPU optional

APIs

  • POST /predict
  • Accepts image upload
  • Returns predicted emotion + confidence score

🔐 9. Privacy & Security

This system:

✔ Does not store user images locally ✔ Logs only label + confidence to Snowflake ✔ Uses Snowflake Secrets (env vars) ✔ Uses HTTPS when deployed ✔ Can anonymize all metadata

📝 10. Future Improvements

  • Upgrade to a deeper CNN or ViT
  • Bias reduction via dataset balancing
  • Add adversarial robustness
  • Add face detection preprocessor
  • Add calibration for confidence scores
  • Improve evaluation suite
  • Add multilingual documentation