Architecture Model
The model is based on ResNet18 as a backbone and fine-tuning only the last convolutional block (layer4). It also use custom classifier for Hidden layer of 256 units with ReLU activation, Dropout layer (0.3) for regularization, and Output layer with 6 units (one per waste category). The dataset is from garythung/trashnet
Model Performance
The model achieves strong performance across all waste categories, with an overall accuracy of 96%. Detailed performance metrics per class:
| Category | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| Cardboard | 0.99 | 0.95 | 0.97 | 161 |
| Glass | 0.92 | 0.98 | 0.95 | 200 |
| Metal | 0.97 | 0.93 | 0.95 | 164 |
| Paper | 0.97 | 0.96 | 0.96 | 238 |
| Plastic | 0.96 | 0.95 | 0.96 | 193 |
| Trash | 0.90 | 0.96 | 0.93 | 55 |