An image classification pipeline that identifies 10 categories of waste from images using transfer learning and deployment-ready model formats.
- Prepared and validated 12,259 images, retaining 12,252 valid samples across 10 waste categories.
- Explored class balance and standardized 2,407 different image resolutions through a consistent preprocessing pipeline.
- Addressed uneven class representation with balanced class weights and strengthened generalization with image augmentation.
- Built an EfficientNetV2S transfer-learning classifier with staged feature extraction and fine-tuning.
- Added reproducible training controls: fixed seeds, checkpointing, early stopping, and learning-rate scheduling.
- Evaluated performance with accuracy, Top-3 accuracy, per-class precision/recall/F1, and a confusion matrix.
- Exported the trained model as TensorFlow SavedModel, TensorFlow.js, and TensorFlow Lite, then verified inference on sample images.
Python · TensorFlow\Keras · EfficientNetV2S · NumPy · Pandas · scikit-learn · Matplotlib · KaggleHub · Jupyter Notebook
- 94.89% test accuracy
- 99.40% Top-3 accuracy
- 0.94 macro F1
- 20.96 MB TensorFlow Lite model
- Verified inference across the 10 target categories.