Deep Learning mini-project: classifying household waste into 10 categories using CNNs and Transfer Learning. Built with TensorFlow/Keras, trained on Kaggle (P100 GPU).
Best model: 93.51% test accuracy, 0.930 macro F1 (ResNet50 fine-tuned with class weights).
Multi-class image classification for automated waste sorting / recycling.
- Input: RGB image resized to 224×224×3
- Output: one of 10 classes — battery, biological, cardboard, clothes, glass, metal, paper, plastic, shoes, trash
- Metrics: accuracy and macro F1 (robust to class imbalance)
Garbage Classification V2 by Suman Kunwar — 12,259 images across 10 classes.
- Significant class imbalance: 4.18× ratio (clothes: 1892, trash: 453)
- Stratified 70/15/15 train/val/test split (proportions preserved to <0.1%)
- Reference paper benchmark: 95.13% (EfficientNetV2S at 384×384)
| # | Model | Test Acc | Macro F1 |
|---|---|---|---|
| 1 | Baseline CNN (from scratch) | 78.61% | 0.780 |
| 2 | ResNet50, frozen base | 92.69% | 0.920 |
| 3 | ResNet50, fine-tuned | 93.29% | 0.925 |
| 4 | EfficientNetV2S, frozen | 92.53% | 0.919 |
| 5 | ResNet50 fine-tuned + class weights | 93.51% | 0.930 |
- Transfer Learning is decisive: +14 pp over a from-scratch CNN, with validation accuracy exceeding the final baseline after a single epoch.
- Class weights solved the minority-class bottleneck: the
trashclass recall was stuck at 0.81 across three architectures; inverse-frequency weighting lifted it to 0.88 while also improving overall accuracy. - Architecture cannot be separated from input resolution: EfficientNetV2S did not beat ResNet50 at 224×224 because it is designed for 384×384 (compound scaling), which also explains the gap to the paper benchmark.
- EDA predictions held up: three confusion pairs predicted from visual inspection (paper/cardboard, glass/metal, plastic/glass) all appeared in the confusion matrices.
trash-classification/
├── notebooks/ # EDA, data pipeline check, and 5 experiment notebooks
├── src/
│ ├── data.py # stratified split, augmentation, class weights (tf.data)
│ ├── model.py # baseline CNN, ResNet50, EfficientNetV2S builders
│ ├── train.py # training loop with EarlyStopping / ReduceLROnPlateau
│ └── utils.py # metrics logging, confusion matrix, plots
├── experiments/
│ ├── results.csv # logged metrics for all experiments
│ └── plots/ # training curves and confusion matrices
└── report/ # final report (PDF)
Local (development / debugging):
conda create -n trash python=3.11 -y
conda activate trash
pip install -r requirements.txtTraining (Kaggle):
Open a notebook from notebooks/, attach the dataset
sumn2u/garbage-classification-v2, set the accelerator to GPU, and run all
cells. Each notebook clones this repo and imports the reusable modules from
src/.
TensorFlow / Keras · scikit-learn · NumPy / pandas · matplotlib / seaborn · Kaggle (P100 GPU)
Code is released under the MIT License (see LICENSE). The dataset is subject to its own license — see the original Kaggle page.