Machine learning model for waste images classification with multi-label class.

We import the dataset from Kaggle, Roboflow, and web scraping in Google Image.

Here is the link to our final dataset: Trashcan Final Dataset

Libraries that we used for preprocessing the images and training the model are:

os
shutil
numpy
sklearn
seaborn
matplotlib
tensorflow
keras

Each images has 2 labels, category and sub-category.

Here are the category labels:

• Organik
• Anorganik
• B3

And here are the sub-category labels:

• Daun
• Kardus
• Makanan Olahan
• Kaleng
• PET
• Tas Plastik Belanja
• Aerosol
• Baterai
• Obat Kapsul

We used transfer learning to train the model with EfficientNetB3V2 as the base model and reached an overall accuracy of 95%.

Here is the link to our final model: Trashcan Final Model

Confusion matrix for category

Confusion matrix for sub-category

The final model was saved in .h5 and the API ran in Flask Python. The Flask app later deployed in cloud using Docker and Cloud Run.

Flask==3.0.3
numpy==1.26.4
tensorflow==2.16.1
Werkzeug==3.0.3
pillow==10.3.0

• Local
  • Clone this repository
  • Make sure to download the model that we provide in here.
  • Run pip install -r requirements.txt command on CMD to install the required libriaries.
  • Run python app.py command
  • Use the URL from the response, https://127.0.0.0:5000/predict to test it in Postman or anywhere else.
• Direct
  • Use the deployed URL, https://trashcan-qnitsxbfya-et.a.run.app/predict directly.