🐕 Dog Breed Classification — CNN from Scratch + ResNet-50 Transfer Learning

🚀 Overview

An image classification pipeline that accepts any user-supplied image and determines: (1) if it's a dog, it identifies the breed from 133 classes, or (2) if it's a human, it identifies the most resembling dog breed. The project builds two CNNs — one from scratch (35% accuracy) and one using ResNet-50 transfer learning (79% accuracy) — combined with OpenCV Haar cascades for face detection and VGG-16 for dog detection.

📊 Results

Model	Test Accuracy	Test Loss
Random Baseline	0.75% (1/133)	—
CNN from Scratch (5 conv layers, 40 epochs)	35%	2.779
ResNet-50 Transfer Learning	79%	0.901

Detection Accuracy:

• Human face detector (Haar cascade): 98% on human images, 83% rejection on dog images
• Dog detector (VGG-16, ImageNet classes 151–268): used for dog/not-dog classification

✨ Key Features

• Full Detection Pipeline — Input any image → detect human (Haar cascade) or dog (VGG-16) → classify breed (ResNet-50) → output breed name or resembling breed
• CNN from Scratch — 5 convolutional layers (3→16→32→64→128→256) with max pooling, dropout, and a single FC layer mapping to 133 breeds
• Transfer Learning — Pre-trained ResNet-50 with frozen feature layers and replaced final FC layer (fc: 2048 → 133), achieving 79% accuracy
• Human Face Detection — OpenCV Haar cascade (haarcascade_frontalface_alt.xml) for detecting human faces in images
• Dog Detection — VGG-16 pre-trained on ImageNet; classes 151–268 correspond to dog breeds, used as a binary dog detector

🧠 Technical Highlights

CNN from Scratch:

Input (3, 224, 224)
  → Conv2d(3→16, 3×3) → ReLU → MaxPool(2×2)     → (16, 112, 112)
  → Conv2d(16→32, 3×3) → ReLU → MaxPool(2×2)     → (32, 56, 56)
  → Conv2d(32→64, 3×3) → ReLU → MaxPool(2×2)     → (64, 28, 28)
  → Conv2d(64→128, 3×3) → ReLU → MaxPool(2×2)    → (128, 14, 14)
  → Conv2d(128→256, 3×3) → ReLU → MaxPool(2×2)   → (256, 7, 7)
  → Flatten → Dropout(0.2) → Linear(256*7*7 → 133)

Transfer Learning (ResNet-50):

• All convolutional layers frozen (pre-trained on ImageNet)
• Final fc layer replaced: Linear(2048 → 133) for 133 dog breeds
• Only the new FC layer is trained, leveraging ResNet-50's learned feature representations

Why This Problem Is Hard:

• Minimal inter-class variation: Brittany vs. Welsh Springer Spaniel look nearly identical
• High intra-class variation: Labradors come in yellow, chocolate, and black
• 133 classes with limited training data per breed

🛠 Tech Stack

Component	Technology
Framework	PyTorch
Transfer Learning	ResNet-50 (ImageNet pre-trained)
Dog Detection	VGG-16 (ImageNet classes 151–268)
Face Detection	OpenCV Haar Cascade
Loss	CrossEntropyLoss
Optimizer	SGD (lr=0.01) for scratch, Adam for transfer
Dataset	8,351 dog images across 133 breeds

🏗 Pipeline Architecture

User-Supplied Image
    │
    ├── OpenCV Haar Cascade → Human detected?
    │   └── Yes → ResNet-50 → "You look like a {breed}!"
    │
    ├── VGG-16 (ImageNet 151-268) → Dog detected?
    │   └── Yes → ResNet-50 → "Predicted breed: {breed}"
    │
    └── Neither → "Error: No human or dog detected"

⚡ Getting Started

git clone https://github.com/jashjain21/CNN-Dog_breed_classification.git
cd CNN-Dog_breed_classification

pip install torch torchvision opencv-python numpy matplotlib

# Download datasets (required)
# Dog images: https://s3-us-west-1.amazonaws.com/udacity-aind/dog-project/dogImages.zip
# Human faces: http://vis-www.cs.umass.edu/lfw/lfw.tgz

jupyter notebook dog_app.ipynb

Test images are provided in upload_images_for_testing/ for quick evaluation.

🔍 What This Project Demonstrates

• Transfer Learning — Replacing the final layer of a pre-trained ResNet-50 and fine-tuning for a domain-specific task, jumping from 35% (scratch) to 79% accuracy
• Multi-Model Pipeline — Chaining Haar cascades, VGG-16, and ResNet-50 into a single inference pipeline that handles different input types (human, dog, other)
• CNN Architecture Design — Building a 5-layer CNN from scratch with progressive channel depth (16→32→64→128→256) and evaluating its limitations
• Computer Vision Fundamentals — Face detection, image classification, data augmentation, and the challenges of fine-grained visual categorization (133 visually similar breeds)

🚧 Limitations / Future Improvements

• 79% Accuracy — While above the 60% threshold, fine-tuning more ResNet layers (not just the FC) or using a larger model (ResNet-152, EfficientNet) could push accuracy higher
• No Data Augmentation on Transfer Model — Adding random rotations, flips, and color jitter during training would improve generalization
• Haar Cascade Limitations — The face detector misclassifies 17% of dog images as containing faces; a deep learning face detector (MTCNN, RetinaFace) would be more robust
• No Top-K Predictions — Only the top-1 breed is returned; showing top-3 with confidence scores would be more useful for similar-looking breeds
• No Web/Mobile Interface — The pipeline runs in a notebook; wrapping it in a Flask/FastAPI app would make it user-facing

Author

Jash Jain : LinkedIn