Awesome

Arduino trash classification TinyML example

SNUCSE 2020 "creative integrated design"

TinyML Team (with LG Electronics)

• Myeonghwan Ahn
• Junseo Koo
• Jaewoo Kim
• LGE mentor Chanseok Kang

LGE lent us

• Arduino Nano 33 BLE Sense
• ArduCAM OV2640
• F/F cables to connect above two

which are necessary for project

Description

TinyML example application for Arduino Nano 33 BLE Sense

demo video

tested environment

• arduino nano 33 BLE Sense + ArduCAM OV2640
• macOS Big Sur (v11.1, 20C69), MBPr 13" mid 2014
• arduino IDE v1.8.13
• arduino library Arduino_TensorFlowLite 2.1.0-ALPHA
• arduino library JPEGDecoder v1.8.0
• arduino library ArduCAM v1.0.0

model

MobileNet v1 25% on 96x96 RGB input

6-way classification

• cardboard
• glass
• metal
• paper
• plastic
• trash

MobileNet v1 model was trained on custom ImageNet/96x96 dataset

and transfer learning into resized_trashnet, original dataset from garythung/trashnet

how to use

1. follow official person detection example instructions
2. git clone https://github.com/lightb0x/arduino_trash_classification.git in directory of your taste
3. mv arduino_trash_classification-master arduino_trash_classification
4. open arduino_trash_classification OR gopen arduino_trash_classification (open in files)
5. double click on arduino_trash_classification.ino

performance

• speed : takes about 930ms per inference (takes 260ms on Raspberry Pi Zero W)
• accuracy : correctly infer plastic on transparent plastic bottle

reason for performance boost

End-to-end int8 quantization make use of CMSIS-NN kernel instead of standard TFLite micro kernel.

refer here for detail.