Waste Classification Using CNN

Classify waste images into 9 classes using the RealWaste dataset (4,752 images), as a college deep learning project.

Started with a baseline CNN, then fine-tuned MobileNetV2 using transfer learning with two-phase training. The winning ideas from the team's experiments were combined into the final configuration, with 5-crop test-time augmentation on top.

92.99% test accuracy and 0.9359 macro-F1, up from 69.7% for the baseline CNN and 84.3% for the early MobileNetV2. The weakest class is Miscellaneous Trash (F1 0.87). Presented and defended live.

Streamlit bugs: CSS leaking out as text, camera Stop-button lag, and a theme toggle crash. Also, a glass tumbler was misclassified because of a domain gap in the dataset.

Transfer learning made the biggest jump, from 69.7% for my baseline CNN to 84.3% with an early MobileNetV2, and combining the team's winning ideas got it to 92.99%. A strong test score doesn't guarantee real-world behaviour: the glass tumbler was misclassified because of a domain gap in the dataset. Wiring the model into Streamlit brought its own bugs, so the app needed debugging too, not just the model.