Apppublic

mariadolores00/facial-emotion-blendshapes

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

🎭 Facial Emotion Recognition — Blendshapes + ML

Real-time facial emotion recognition using MediaPipe Blendshapes as an anonymized facial representation, combined with a Machine Learning model trained on that data.

🔍 How it works

1.Captures image from webcam
2.MediaPipe extracts 52 blendshapes (numerical values describing facial geometry)
3.A ML model classifies the emotion into: Neutral · Happy · Sad · Angry
4.Displays the prediction with per-class probabilities

🔒 Privacy by design

Blendshapes are abstract mathematical vectors. No images are stored, and it is not possible to reconstruct the original face from the data. This was a deliberate architectural decision to embed privacy from the ground up.

🏗️ Full pipeline

This project is part of an end-to-end pipeline built with two tools:

Tool	Technology	Purpose
Capture app	Flask + MediaPipe (JS)	Captures emotion-labeled blendshapes and generates anonymized CSV datasets
Training app	PyQt5 + TensorFlow + scikit-learn	Trains KNN, SVM, Decision Tree or Neural Network models on the datasets
This demo	Gradio + MediaPipe + TF	Real-time inference with a trained model

🧠 Supported models

—K-Nearest Neighbors (KNN)
—Support Vector Machine (SVM)
—Decision Tree
—Neural Network (TensorFlow/Keras) — .h5

Note: This demo accepts .h5 neural network models. For KNN/SVM/DT, the full pipeline uses joblib .pkl.

🛠️ Tech stack

Python · MediaPipe · TensorFlow/Keras · scikit-learn · Gradio · OpenCV · Flask

👩‍💻 Author

María Dolores Belchí Martínez — Electronics Engineer | Computer Vision | AI Engineer LinkedIn · GitHub