Flower Species Classification using CNN (Oxford 102)

Telling a rose from a camellia is easy; telling 102 flower species apart — many differing only in petal texture or stamen shape — is a genuinely hard vision problem called fine-grained classification. The Oxford 102 Flowers dataset (8,189 photographs, 102 categories, some with as few as 40 images) is the classic benchmark for it, and it is deliberately awkward: too small to train a deep network from scratch, too fine-grained for a shallow one. The right answer is transfer learning, and this project does it the rigorous way — first a linear probe on a frozen ResNet-50 backbone, then fine-tuning of the top residual blocks with a low learning rate and discriminative scheduling. A web demo makes the result tangible: pick a bloom photo and see the ranked top-5 species with confidence scores, exactly like a plant-identification app.

1. The Oxford 102 Flowers dataset (8,189 images, 102 classes, 40–258 images per class) is loaded with the official train/validation/test splits; train+validation are combined for fitting.
2. Images are resized to 224×224×3 and normalized with ImageNet statistics; augmentation applies random crops, horizontal flips and color jitter.
3. Phase 1 (linear probe): the ResNet-50 backbone stays frozen while only the new classification head trains, establishing a baseline.
4. Phase 2 (fine-tuning): the top residual blocks are unfrozen and trained with a 10× lower learning rate plus early stopping on validation loss.
5. The held-out test set is evaluated once: top-1/top-5 accuracy, macro F1 and a confusion analysis over the most-mixed species pairs.
6. In the web demo, a bloom photo passes through identical preprocessing and the saved model, and the app renders the ranked top-5 species with confidence bars.

Tech stack:

• Python 3, PyTorch (torchvision ResNet-50)
• Transfer learning (ImageNet weights)
• NumPy, PIL (preprocessing)
• Matplotlib, scikit-learn (evaluation)
• Jupyter Notebook (training)
• HTML/CSS/JavaScript (web demo)
• Oxford 102 Flowers dataset (VGG, Oxford)

Dataset & model details

• Dataset: Oxford 102 Flowers — Visual Geometry Group, University of Oxford (Nilsback & Zisserman). 8,189 photographs across 102 flower categories (40–258 images per class); official split ≈ 1,020 train / 1,020 validation / 6,149 test.
• Task: 102-class fine-grained image classification; input = 224×224×3 image, output = probability distribution over 102 species.
• Model: ResNet-50 (ImageNet pretrained) → GlobalAveragePooling → Dropout(0.4) → Dense(512, ReLU) → Dense(102, softmax); two-phase training: frozen-backbone linear probe, then fine-tune top blocks at low LR.
• Metrics: Top-1 accuracy (design target ≈ 94%), top-5 accuracy (design target ≈ 99%), macro F1, confusable-species analysis. No accuracy is claimed as measured until the training run is executed for the order.

• [102-species classifier] ResNet-50 fine-tuned on Oxford 102 Flowers — identifies species from a single bloom photograph with top-5 ranking.
• [Live web demo] Select sample blooms (rose, sunflower, daisy, tulip, orchid, lily) and get the predicted species, Latin name and top-5 confidence bars.
• [Two-phase transfer learning] Documented linear-probe-then-fine-tune schedule with layer-wise learning rates, exactly as the literature recommends.
• [Full training notebook] Dataset loading, 224×224 preprocessing, augmentation (crops, flips, color jitter), training loops and evaluation in one reproducible notebook.
• [Top-1 / top-5 evaluation] Accuracy plus macro F1 and a confusable-species analysis (which species the model mixes up, and why).
• [Training curves & error gallery] Validation curves and a misclassified-samples gallery with interpretation notes for the report.
• [Exported fine-tuned model] Saved weights plus preprocessing code, so the demo runs the real network without retraining.

• Complete training & evaluation Jupyter notebook (PyTorch)
• Fine-tuned ResNet-50 weights with preprocessing code
• Web demo wired to the trained model (sample blooms + top-5 ranking)
• Training curves, error gallery and confusable-species analysis
• Project report PDF (background, transfer-learning theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (fine-grained classification, fine-tuning, top-k metrics)

• 102 species from one dataset — garden varieties outside these classes will be force-fit to the nearest species.
• ≈ 94% top-1 / 99% top-5 are design targets for the training run, stated honestly — the report documents the actual achieved figures after training.
• Classes with only ~40 training images remain the weakest; the report quantifies this per-class.
• The model expects a reasonably centered bloom photo; wide garden scenes with tiny flowers are out of scope.
• This is a teaching build, not a botanical reference — no claim is made about taxonomic authority.