ML - Facemask Detection i.MX 8M Plus NPU Demo

This blog post will detail how Boundary Devices created its Facemask Detection app for the i.MX 8M Plus-based Nitrogen8MP and its NPU! It was done as part of our i.MX 8M Plus Machine Learning webinar which you can see here:

• Machine Learning 101

Facemask Detection app details

History

We decided to create a demo leveraging the i.MX 8M Plus NPU when COVID was spreading worldwide. So it became obvious that an application detecting whether or not a person is wearing a mask was a great example. Not only does it show how machine learning can be used to track and recognize a face, a model can be created to learn if the detected face has a specific attribute, like a mask.

Hardware setup

For this demo, we used our Nitrogen8M Plus Evaluation Kit Bundle which includes:

• Nitrogen8MP SoM
• Nitrogen8MP EVK Carrier
• Basler daA3840 8MP camera
• BD070LIC2 7" 1280x800 display

Software details

Overall architecture

After some investigation, we decided to use the following:

• OS: Android 10 (BSP 2.5.0)
  • Pre-built images available here
  • Plenty of examples apps available
  • NNAPI supported by NXP
• Inference Engine: TensorFlow Lite
  • Using existing example to build demo
  • Customized model for face mask detection

Leveraging existing projects

This application was possible thanks to several other projects: Credit: Esteban Uri (Medium article)

• Dataset from Prajna Bhandary
  • Cleverly generated “with mask” images
  • github.com/prajnasb/observations
• 1st model generation from Adrian Rosebrock
  • Face Mask Detector with OpenCV, Keras/TensorFlow
• Android app created by Esteban Uri
  • Real time face mask detection in Android
  • Adaptation to TensorFlow Lite (float32)
• Boundary Devices additions!
  • Quantized model (int8) to run on NPU
  • Allow landscape mode

Demo

Here is a snippet of our webinar showing the Facemask Detection app which can be downloaded here.