Butterfly Species Classification
4-phase multi-model pipeline for 75-species classification. Vanilla CNN → pretrained TL → hybrid parallel/sequential → multi-loss auxiliary heads. Grad-CAM confirms wing-pattern focus. t-SNE shows inter-species clustering.
5,199 train / 1,300 val, 75 butterfly species
4-phase: vanilla CNN → pretrained TL → hybrid parallel → multi-loss auxiliary heads
Advanced 4-phase training strategy for fine-grained 75-class butterfly species classification.
Dataset
- ▸5,199 training + 1,300 validation images, 75 species
- ▸Natural backgrounds, varying lighting, occlusion
- ▸Augmentation: rotation ±30°, flip, zoom 0.2, brightness/contrast ±20%
4-Phase Progressive Training
| Phase | Architecture | Key Feature |
|---|---|---|
| 1 | Vanilla CNN | Standard conv blocks, augmentation baseline |
| 2 | Pretrained CNNs | VGG16/ResNet frozen → fine-tuned |
| 3 | Hybrid parallel/sequential | Multi-scale parallel branches merged |
| 4 | Multi-loss auxiliary | Auxiliary heads at intermediate layers |
Multi-Loss Training Auxiliary classification heads at intermediate layers provide gradient signals deeper into the network. For 75-class fine-grained recognition, this reduces the vanishing gradient problem and forces intermediate features to be semantically meaningful.
Hybrid Architecture Parallel branches capture features at different receptive field sizes simultaneously — critical since butterfly wing patterns exist at both fine (vein structure) and coarse (color patch) scales.
Interpretability
- ▸Grad-CAM: Model correctly focuses on wing patterns and coloration, not background
- ▸t-SNE: 2D feature space shows clear inter-species clustering despite visual similarity
Fine-Grained Challenge 75-class butterfly ID is hard because: (1) high intra-class variation (same species at different angles), (2) high inter-class similarity (similar wing patterns), (3) natural backgrounds. The multi-scale + multi-loss strategy directly addresses all three.