Image and Audio
Recognition with Edge
Impulse
By John Staveley MVP
23/05/2023
https://uk.linkedin.com/in/johnstaveley/
@johnstaveley
Overview
 Creating projects in EdgeImpulse
 Linking a device
 Capturing and classifying data
 Choosing and refining a model
 Deploying and using an image recognition model
 Audio recognition in EdgeImpulse
 Deploying the audio recognition model
Creating
a project
DEMO
Set data
source
https://docs.edgeimpulse.co
m/docs/development-
platforms/fully-supported-
development-boards
Use your phone
 Use your mobile phone:
Add device – scan QR code for mobile phone
 Opens in web page - no app required
 Take 30-40 pictures from different angles
 Split into train and test data
Set labels
Create model
Transfer learning
Generate features
Generate model
Generate model
 Wait 20 minutes
https://stephenallwright.com/interpret-f1-score/
Training length v performance
 Training cycles : F1 Score : Time
 60 : 0% : 5 mins
 130 : 26.7% : 10 mins
 250 : 57.1% : 20 mins
Deploying the model
 Scan the QR code
Other Deployments
 PC
 WebAssembly
 Arduino
 Raspberry Pi
 Jetson Nano
 Export model in Tensorflow
 using Microsoft.ML.Tensorflow
Audio detection using Edge Impulse
 Data collection, split 80/20%
 Generate features
 Select and train model
 Deployment
 Download model
 Upload to Nicla voice with Arduino sketch
Conclusion
 Data acquisition, model creation and deployment are easy in EdgeImpulse
 Train the model for as long as possible
 Resultant model can be used in a number of platforms
Resources
 https://www.edgeimpulse.com
 https://www.youtube.com/watch?v=dY3OSiJyne0&t=105s
 https://docs.arduino.cc/tutorials/nicla-voice/getting-started-ml
Any Questions?
@johnstaveley
Slides:
https://www.slideshare.net/johnstaveley/
Question for the pub:
 What will data science as a profession look like in 10 years time?
 Will it all be just running the tools?

Image and Audio Detection using Edge Impulse