Waste Classification using CNN

Point a camera at a piece of waste and the model tells you what it is.
9 categories, 92.99% test accuracy, photo upload and live webcam.

---

Overview

A college deep learning project built by a team of 7. We trained a baseline CNN and a transfer learning model on the RealWaste dataset (4,752 images, 9 classes), improved the better model through four separate experiments, and wrapped the result in a Streamlit app with live webcam classification.

The project was presented and defended live.

Results

Model	Test accuracy	Macro-F1
Baseline CNN (4 conv blocks, trained from scratch)	78.54%	78.61%
MobileNetV2, fine-tuned + improved (final)	92.99%	0.9359

• Strongest classes: Cardboard, Textile Trash and Vegetation (F1 about 0.96 to 0.98).
• Weakest class: Miscellaneous Trash (F1 0.8707).
• Most common mistake: Plastic and Metal get mixed up, though far less than in earlier runs.
• Before test-time augmentation, two full training runs landed at 92.85% and 92.43%. We did not seed PyTorch, so expect around ±0.4% between runs.

How it works

Data. RealWaste, 4,752 images across 9 classes. Preprocessing keeps the aspect ratio when resizing, normalizes separately for the baseline and the pretrained model, and checks for corrupted files.

Models.

1. Baseline CNN: 4 conv blocks (32 / 64 / 128 / 256 filters), a dense layer, dropout.
2. MobileNetV2 (ImageNet pretrained), trained in two phases: first only the new classification head, then the last layers unfrozen at a low learning rate.

Model improvement. Four teammates each tested one idea on MobileNetV2, and we combined the winners:

Idea	What we found
How many layers to unfreeze	Deeper unfreezing helped; we unfreeze whole blocks (features[-11:]) so Conv and BatchNorm layers stay paired
Optimizer, LR schedule, dropout	Tuned values with ReduceLROnPlateau gave a clear gain
Loss function	Label Smoothing beat Focal Loss
Test-time augmentation	5-crop averaging at inference raised test accuracy by about 1 point

The app

Built with Streamlit, with three tabs:

• Classify Image: upload a photo and see a confidence bar for every class.
• Live Camera: webcam classification through OpenCV.
• Model Specs: details about the model.

There's also a light and dark theme. The checkpoint path is a text field in the app, so it runs on any laptop without editing code.

Getting started

git clone <YOUR-REPO-LINK>
cd <repo-folder>
pip install torch torchvision streamlit opencv-python pillow
streamlit run app.py

Put the trained checkpoint (task7_combined.pth) next to app.py, or paste its path into the app's sidebar.

Known limitations

• The model learned from RealWaste, which is photos of waste at a sorting facility. Clean product photos (a glass tumbler on a white background, for example) can be misclassified. That's a domain gap in the data, not a bug in the app.
• The webcam tab classifies short bursts of frames, not a continuous video stream. streamlit-webrtc would be the upgrade.
• Plastic, Metal and Glass remain the hardest classes to separate.

Team

Team of 7, built at Government Polytechnic Mumbai.

Yashmit Katekar (team lead)	Dataset and preprocessing, CNN research, baseline model, loss function experiment, model improvement
Mukta (team lead)	Transfer learning architecture research, unfreeze depth experiment
Shreeyans	Optimizer, scheduler and dropout tuning
Rudra	Test-time augmentation

Acknowledgements

Dataset: RealWaste, from the UCI Machine Learning Repository.