Dog Breed Identification from Photos using CNN

Telling a Beagle from a Foxhound is far harder than telling a dog from a cat: fine-grained classes differ only in subtle cues like ear set, muzzle shape and coat texture. Training a deep network from scratch for 120 such classes would need enormous data and compute. Transfer learning is the professional answer — reuse visual features learned from ImageNet and retrain only the classifier head. The Stanford Dogs dataset (Khosla et al., Stanford Vision Lab) provides 20,580 annotated images across 120 breeds with a standard train/test split, making it the established benchmark. This project implements the full pipeline: preprocessing, two-stage fine-tuning, augmentation for per-breed data scarcity, honest test-split evaluation, and a photo-upload demo showing top-3 breeds with breed facts.

1. The photo is resized to 224x224 and normalized with ImageNet mean/std statistics.
2. The frozen ResNet-50 backbone converts it into a 2048-dimensional feature vector of fur texture, ear geometry and proportions.
3. The classification head (global average pooling → Dense(512) → dropout → Dense(120)) maps features to breed logits.
4. Softmax converts logits to 120 probabilities; the demo reports the top three.
5. Stage-one training fits only the head (Adam, lr 1e-3); stage two fine-tunes the last block at lr 1e-5.
6. Final evaluation runs once on the held-out test split for top-1/top-5 accuracy.

Tech stack:

• Python 3.10, PyTorch or TensorFlow/Keras
• ResNet-50 (ImageNet pre-trained)
• Stanford Dogs dataset (Khosla et al.)
• Albumentations / torchvision transforms
• Flask (demo backend)
• HTML/CSS/JavaScript (demo UI)
• Git

Dataset & Model Details

• Dataset: Stanford Dogs — 20,580 images, 120 breeds (~150 per breed), with bounding-box annotations; source: Aditya Khosla et al., Stanford Vision Lab. Standard split: 12,000 train / 8,580 test.
• Model input: 224x224x3 normalized image. Output: 120-way softmax (breed probabilities).
• Architecture: ResNet-50 backbone (ImageNet init) → GlobalAvgPool → Dense(512) → Dropout → Dense(120).
• Metric: top-1 / top-5 accuracy on the test split, measured after training; per-breed precision analysis included. No accuracy figure is claimed before the training run.

• [120-breed classifier] ResNet-50 backbone with a retrained classification head covering all 120 Stanford Dogs breeds.
• [Two-stage fine-tuning] Head-only training first, then the last residual block unfrozen at a low learning rate with cosine decay.
• [Top-3 predictions with breed cards] Each prediction shows probabilities plus group, size, coat and temperament notes per breed.
• [Augmentation pipeline] Random crop, flip, color jitter and ±15° rotation counter the small per-breed sample sizes.
• [Per-breed error analysis] Confusion analysis highlights commonly confused pairs (e.g. Malamute vs. Husky) with measured numbers.
• [Photo-upload web demo] Upload any dog photo; get breed probabilities in under a second on CPU.
• [Honest evaluation] Top-1/top-5 accuracy measured once on the 8,580-image held-out test split — reported as measured.

• Python training notebook (two-stage transfer learning, augmentation, evaluation)
• Fine-tuned model weights
• Test-split evaluation: top-1/top-5 accuracy + per-breed error analysis from the actual run
• Photo-upload web demo with breed information cards
• Project report PDF (methodology, ablations, measured results)
• PPT presentation for final review
• Viva Q&A preparation document