Building a Facial Recognition Doorbell at the Edge
Applied transfer-learning to facial image encoding model on edge device to enhance performance and user privacy.
Description
The project is focused on applying transfer learning to a pre-trained neural network at the edge in order to build a customized facial recognition system that can be used for smart doorbells. The benefit of transfer learning at the edge is that user-specific data and models won’t need to be uploaded to the cloud since uploading sensitive user data to the cloud could be a security concern.
We used FaceNet, a neural network developed in 2015 by reserachers at Google for face recognition as our base model and performed transfer learning on Nvidia Jetson TX2. The result is an exciting proof of concept that demonstrates it is possible to avoid cloud training for detecting known user identities, allowing facial recognition to be deployed securely and avoiding the risks of transmitting user identities to the cloud.
Outcome
The resulting model achieved over 99% accuracy on recognizing user identities even for images with high degrees of light or shadow.
Demo
Techniques
- supervised learning (classification) with SVM
- facial recognition
- image embedding with FaceNet
- transfer learning
- containerization
- model deployment at the edge
Tools
- Keras
- sckit-learn
- Docker
- Nvidia TX2
- Linux
More Information
More information can be found at the following links: