This project focuses on recognizing micro facial expressions using Convolutional Neural Networks (CNN). Micro-expressions are subtle, involuntary facial expressions that occur briefly and reveal genuine emotions. Detecting them can be crucial in areas like security, psychology, and suspect interrogation. The model is trained and evaluated using two benchmark datasets: CASME II (for micro-expressions) and FER-2013 (for facial expressions).Let's embark on this exciting journey!
CNN-based architecture for emotion classification
Trained on both macro and micro-expression datasets
Handles subtle facial movements in short image sequences
Comparative analysis using different datasets
Results visualization included
The CNN model consists of:
Input Layer – Preprocessed grayscale facial images
Convolutional Layers – Feature extraction with ReLU activation
Max Pooling Layers – Dimensionality reduction
Flatten Layer – Transforms data for dense layers
Dense Layers – For final classification
Output Layer – Softmax activation to classify emotion categories
CNN Architecture Diagram:
- CASME II Focuses on spontaneous micro-expressions
Contains high frame rate facial sequences
Used for training a model to recognize subtle expressions
Training Accuracy: 92%
- FER-2013 Publicly available dataset with 35,000+ labeled facial expression images
Contains 7 emotion categories
Training Accuracy: 73%
Dataset Link: https://www.kaggle.com/datasets/msambare/fer2013
Dataset Training Accuracy CASME II 92% FER-2013 73%
CASME II showed superior performance due to focused and high-quality micro-expression data
FER-2013 provided good generalization for common facial expressions
https://www.kaggle.com/datasets/msambare/fer2013
https://paperswithcode.com/sota/micro-expression-recognition-on-casme-ii-1
https://github.com/yuxinhe/CASME2-Micro-Expression-Database-SVM
https://ieeexplore.ieee.org/abstract/document/10820184
Muqadas Ejaz
BS Computer Science (AI Specialization)
Machine Learning & Computer Vision Enthusiast
📫 Connect with me on LinkedIn
🌐 GitHub: github.com/muqadasejaz