WasteAI v1.0.0 — Initial Public Release

WasteAI is an academic deep-learning project for image-based waste classification. This release brings together four trained CNN architectures, a public Streamlit application, Hugging Face model hosting, and model evaluation visualizations.

Features

• Classification into six categories: Cardboard, Glass, Metal, Paper, Plastic and Trash.

• Four selectable deep-learning models:

  • Custom CNN

  • MobileNetV2

  • VGG16

  • VGG19

• Image upload and prediction through a Streamlit interface.

• Model performance comparisons and training analytics.

• Confusion matrices, training curves and evaluation reports.

• Hugging Face-based model loading.

• Local development and deployment documentation.

Model evaluation

Recorded held-out test-set results:

Model | Accuracy | Macro F1 -- | -- | -- VGG19 | 91.29% | 89.93% VGG16 | 88.65% | 87.69% MobileNetV2 | 87.07% | 86.01% Custom CNN | 73.09% | 69.62%

The experiments use 2,526 images: 1,768 training, 379 validation and 379 testing images.

Project links

• Live application: https://wasteai-atchyuteswar.streamlit.app/

• GitHub repository: https://github.com/Atchyuteswar/WasteAI

• Hugging Face models: https://huggingface.co/Atchyuteswar/wasteai-models

Important limitations

WasteAI is an educational research prototype, not a certified waste-management or recycling system. Test-set performance does not guarantee accuracy on unseen real-world waste. Model scores are not necessarily calibrated probabilities, and disposal guidance should be checked against local recycling rules.

The original dataset and model weights are hosted separately. Refer to the repository documentation for setup, reproduction and third-party licensing information.

Thank you for exploring WasteAI!