Image Classification on TinyImageNet-100

Comparative study of computer vision techniques: from traditional feature extraction to deep convolutional networks.

Project Overview

Dataset: TinyImageNet-100 (100 classes, 64x64 images) Phases:

1. Traditional CV: SIFT/ORB + Bag of Words + Fisher Vector + SVM
2. Deep Learning (FCNN): Fully Connected Networks with Keras
3. PyTorch Implementation: FCNN with data augmentation and LR scheduling
4. Advanced CNN: Convolutional architectures

---

Key Results Summary

Model	Approach	Key Features	Status
ORB + BoW	Traditional	Bag of Words (k=100), Linear SVC	Baseline
ORB + Fisher	Traditional	GMM (64 components), Fisher Vector	Improved
SIFT + BoW	Traditional	RBF SVM (C=10, γ=0.01)	Robust
FCNN (Keras)	Deep Learning	64x64 flatten, Adam optimizer	Entry DL
FCNN + BN/Dropout	Deep Learning	Batch Norm, Dropout(0.25), Class weights	Regularized
Phase 3 FCNN	Deep Learning	Augmentation, ReduceLROnPlateau	Best FCNN
Phase 4 CNN	Deep Learning	Convolutional Layers	Best Overall

---

Visualizations

1. Traditional Computer Vision Results

ORB + Bag of Words (Linear SVC)

ORB + Fisher Vector (GMM-64)

SIFT + Bag of Words (RBF SVM)

---

2. Deep Learning: Fully Connected Networks (Keras)

Baseline FCNN (64x64 Flatten)

Architecture: Flatten → Dense layers with Adam optimizer

Improved FCNN (BatchNorm + Dropout + Class Weights)

Improvements: Batch Normalization, Dropout(0.25), Validation split, Class weight balancing

---

3. PyTorch Implementation (Phase 3)

FCNN with Data Augmentation & LR Scheduling

Key Techniques:

• Data Augmentation (rotation, flip, color jitter)
• ReduceLROnPlateau scheduling
• 128x128 Grayscale conversion variant

Learning Rate Scheduling Visualization

---

4. Advanced CNN (Phase 4) ⭐ Best Performance

Architecture: Convolutional layers with advanced regularization

---

Detailed Metrics

Detailed classification reports (precision, recall, F1-score per class) are available in:

• results/metrics/ - Text files with full scikit-learn classification reports

Repository Structure

Key Findings

1. Traditional vs Deep Learning: Transition from hand-crafted features (SIFT/ORB) to learned representations shows significant improvement in class separation
2. Regularization Impact: Batch Normalization and Dropout reduced overfitting substantially in FCNN models
3. Data Augmentation: Phase 3 augmentation strategies improved generalization on limited training data
4. CNN Superiority: Convolutional architectures (Phase 4) outperformed fully connected networks by capturing spatial hierarchies

Technologies Used

• Traditional CV: OpenCV (SIFT, ORB), Scikit-learn (SVM, GMM, Fisher Vector)
• Deep Learning: TensorFlow/Keras, PyTorch
• Visualization: Matplotlib, Seaborn
• Data: TinyImageNet-100 (100 classes)

---

Note: Dataset and trained model weights (>100MB) are excluded from this repository per GitHub size limits.