Dog Breed Classification using Transfer Learning

Telling a Siberian Husky from an Alaskan Malamute is hard for people and harder for models: 120 breeds differ in fine details of coat, ears and muzzle that shallow features cannot separate. Training a deep 120-way classifier from scratch would demand enormous labeled data and student-unfriendly hardware — so this build uses transfer learning. A ResNet (He et al., CVPR 2016), pretrained on ImageNet, already knows edges, textures and object parts; its residual skip connections let gradients flow through very deep networks, which is what captures the fine detail separating breeds. The training notebook fine-tunes it on the Stanford Dogs dataset (Khosla et al. 2011) — 20,580 annotated images across 120 breeds — with augmentation and a documented train/validation split. Each run logs accuracy and loss curves and computes a validation confusion matrix with a per-breed breakdown naming the confused pairs, so the report's metrics come from the run itself (design target ~85%+). A demo app serves top-3 predictions with confidence scores and Grad-CAM overlays.

Dataset & model:
Dataset name: Stanford Dogs (Khosla et al., CVPR Workshop on Fine-Grained Visual Categorization, 2011). Source: the public release — 20,580 annotated images across 120 dog breeds. Task: fine-grained image classification. Classes: 120 dog breeds (e.g. Siberian Husky, Alaskan Malamute). A seeded train/validation split is documented in the report.
Model: ResNet with residual connections (He et al., "Deep Residual Learning for Image Recognition", CVPR 2016), fine-tuned from ImageNet-pretrained weights with a replaced 120-way classifier head.
Input: a dog photo (JPEG/PNG), auto-resized and normalized. Prediction: softmax scores over the 120 breeds. Output: top-3 predicted breeds with confidence scores, plus Grad-CAM heatmaps.
Evaluation metrics: validation accuracy, accuracy/loss curves, and a confusion matrix with per-breed breakdown on the validation split — all computed by the training notebook during the build. Design target: ~85%+ validation accuracy — a target for the training run, not a pre-claimed score.

Working:

1. Dataset: Stanford Dogs images (20,580 across 120 breeds) are organized by breed label and split into train and validation sets with a fixed seed for reproducibility.
2. Preprocessing: training images pass through augmentation — random resized crops, horizontal flips, color jitter, rotations — while validation images are only resized and normalized.
3. Training phase: an ImageNet-pretrained ResNet is fine-tuned in two stages — the replaced 120-way classifier head is trained first, then deeper layers are unfrozen with a low learning rate.
4. Logging: each epoch logs training and validation accuracy and loss; the best checkpoint is saved by validation accuracy.
5. Evaluation phase: a confusion matrix with a per-breed accuracy breakdown is computed on the validation split during the build — the report's metrics come from this run, against the design target of ~85%+ validation accuracy.
6. Inference phase (demo): the app loads the best weights at startup; an uploaded photo is preprocessed identically to validation images and passed through the model; softmax scores are ranked into top-3 breed predictions with confidence bars, and Grad-CAM overlays show the attended regions.

Specifications:
Model | ResNet with residual connections (He et al., 2016), fine-tuned from ImageNet weights
Dataset | Stanford Dogs — 20,580 images across 120 breeds (Khosla et al. 2011)
Split | Seeded train/validation split, documented in the report
Metric | Validation accuracy, loss curves and confusion matrix with per-breed breakdown, computed during the build; design target ~85%+
Augmentation | Random resized crops, flips, color jitter, rotations on training images
Input | Dog photos (JPEG, PNG), auto-resized and normalized
Output | Top-3 breed predictions with confidence scores, plus Grad-CAM overlays
Demo app | Flask web app with upload UI and prediction display
Weights | .pth file shipped with the build, plus Grad-CAM visualization support

ResNet Breed Classifier [implemented] — ImageNet-pretrained ResNet fine-tuned on 120 dog breeds, with a replaceable 120-way classifier head.
Two-Stage Fine-Tuning [implemented] — Classifier head trained first, then deeper layers unfrozen at a low learning rate for stable convergence.
Training Notebook with Full Logging [implemented] — Accuracy and loss curves every epoch, plus a validation confusion matrix computed during the build.
Augmentation Pipeline [implemented] — Random crops, flips, color jitter and rotations on training images.
Demo App with Top-3 Predictions [implemented] — Upload a dog photo and get the top-3 predicted breeds with confidence scores and bars; the confidence threshold and top-k display are configurable in the demo config.
Grad-CAM Visualizations [implemented] — Heatmaps showing which regions drove each prediction.
Per-Breed Accuracy Breakdown [implemented] — The confusion matrix names the most-confused breed pairs explicitly, for report error analysis.
Exportable .pth Weights [implemented] — The checkpoint suits deployment-style demos.

Complete source code: data pipeline, training, evaluation, Flask demo app
Trained ResNet weights (.pth)
Dataset download and split-configuration scripts
Project report PDF: background, dataset analysis, methodology, evaluation, error analysis
PPT presentation for final review
Viva Q&A preparation document: residual connections, transfer learning, fine-grained classification, evaluation metrics
Setup guide: environment, dependencies, GPU guidance for training

Fine-grained breeds are the main error source — visually near-identical breeds (e.g. husky vs malamute) confuse the model, as the confusion matrix shows.
The dataset skews toward well-framed photos of single dogs; occluded, multi-dog or low-light photos classify worse.
The classifier is closed-set over the 120 Stanford Dogs breeds; mixed-breed or off-list dogs get the nearest-looking label.
The ~85%+ figure is a design target for the training run, not a pre-measured claim.
This is an educational prototype, not a certified breed-identification service — it must not drive pedigree, insurance or sale decisions.