2025 • Project
Animal and Plant Classifier
Transfer-learning pipelines (VGG16, EfficientNetB0/V2S/V2L) for 10-class animal/plant recognition in TensorFlow.
- TensorFlow
- Keras
- EfficientNetV2
- EfficientNetB0
- VGG16
- Adam
- AdamW
- CNN
- Python
Problem
Build a robust 10-class image classifier for mixed animal and plant categories from foldered images.
Approach
- Data prep: folder split with splitfolders (70/15/15 train/val/test), tf.keras.image_dataset_from_directory, BATCH=32.
- Preprocessing: images at 480×480 then resized to 224×224 in-graph; light aug (RandomFlip, RandomRotation).
- EfficientNet variants: frozen backbones with GAP + dense head — B0, V2S, and V2L; added Dropout in later runs.
- Optimisation: Adam; later AdamW with EarlyStopping + ReduceLROnPlateau.
- Evaluation: top-1 accuracy on held-out test set + per-class accuracy from confusion matrix; training curves logged.
Outcomes
- Best: EfficientNetV2L + Dropout + AdamW + callbacks — 92.13% top-1; ~0.921 macro per-class accuracy.
- 100% grade (HD) for assignment; top 10% of kaggle leaderboard.