CIFAR-100 Image Classification with CNNs

01Project Overview

Building and evaluating convolutional neural networks for fine-grained image classification

This project investigates CNN-based approaches for classifying images from CIFAR-100 — a challenging benchmark with 100 fine-grained categories and only 500 training images per class. We explore models trained from scratch, evaluate robustness to Gaussian noise, implement noise-augmented training for improved resilience, and compare against VGG-16 transfer learning.

69.1%

Best Clean Test Accuracy (CNN)

47.3%

Best Noisy Accuracy (Noise-Aug)

63.3%

VGG-16 Transfer Learning

~2.4M

CNN Parameters

02Dataset & Preprocessing

CIFAR-100: 60,000 32×32 color images across 100 classes

Data Split

Training: 45,000 images
Validation: 5,000 images
Test: 10,000 images

Augmentation & Normalization

Channel-wise normalization (μ, σ per channel)
RandomCrop(32, padding=4)
RandomHorizontalFlip
Batch size: 128

03CNN from Scratch

A moderate-depth CNN with progressive regularization

Block 1

Conv(3→64) · BN · ReLU
Conv(64→64) · BN · ReLU
MaxPool · Dropout(0.2)

Block 2

Conv(64→128) · BN · ReLU
Conv(128→128) · BN · ReLU
MaxPool · Dropout(0.3)

Block 3

Conv(128→256) · BN · ReLU
Conv(256→256) · BN · ReLU
MaxPool · Dropout(0.4)

Classifier

AdaptiveAvgPool(1,1)
FC(256→512) · ReLU
Dropout(0.5) · FC(512→100)

Training Configuration

Optimizer: Adam (lr=1e-3, weight_decay=5e-4)
Scheduler: CosineAnnealingLR (T_max=100)
Loss: CrossEntropyLoss
Epochs: 100
Init: Kaiming He (fan_out, ReLU)

79.5%

Training Accuracy

69.0%

Validation Accuracy

69.1%

Test Accuracy

Training and validation loss/accuracy curves over 100 epochs. Cosine annealing provides smooth convergence with a ~10% train-val gap indicating moderate overfitting.

04Noise Robustness

Evaluating and improving resilience to Gaussian noise (σ²=0.05)

Visual comparison of clean CIFAR-100 images and their noisy counterparts with additive Gaussian noise (σ²=0.05). The noise severely corrupts the low-resolution 32×32 images.

Noise Model

Additive zero-mean Gaussian noise: x_noisy = clamp(x + ε, 0, 1) where ε ~ N(0, 0.05)

Standard CNN under Noise

Condition	Accuracy	Drop
Clean Test	69.1%	—
Noisy Test	1.2%	−67.8%

Finding: The model trained exclusively on clean data collapses to near-random performance under noise — features learned are not noise-invariant.

Noise-Augmented Training

During training, 30% of samples are corrupted with the same Gaussian noise. The model learns noise-invariant features while preserving clean accuracy.

Condition	Standard CNN	Noise-Augmented CNN
Clean Test	69.1%	64.9%
Noisy Test	1.2%	47.3%
Accuracy Drop	67.8%	17.6%

Noise-augmented CNN training curves. The train-val gap is narrower than the clean-only model, showing better generalization due to the regularization effect of noise injection.

05Transfer Learning with VGG-16

Leveraging ImageNet-pretrained features for CIFAR-100

Approach

Load pretrained VGG-16 (ImageNet weights)
Resize CIFAR-100 images: 32×32 → 224×224
Freeze all convolutional layers (feature extractor)
Extract 512-d features via AdaptiveAvgPool
Train lightweight MLP classifier: FC(512→256) → ReLU → Dropout(0.5) → FC(256→100)

63.3%

Clean Test Accuracy

1.5%

Noisy Test Accuracy

~157K

Trainable MLP Parameters

Observation: VGG-16 transfer learning achieves competitive accuracy (63.3%) with only 157K trainable parameters, but is equally vulnerable to noise (1.5% noisy accuracy), suggesting ImageNet features are not inherently noise-robust.

06Comprehensive Comparison

All experiments side-by-side

Model	Clean Acc	Noisy Acc	Drop	Params
bestCNN (scratch)	69.1%	1.2%	67.8%	~2.4M
robustCNN (noise-aug)	64.9%	47.3%	17.6%	~2.4M
transferVGG-16 + MLP	63.3%	1.5%	61.8%	138M + 157K

Clean vs noisy test accuracy and robustness comparison across all models. Noise-augmented training dramatically reduces accuracy drop under noise.

Key Takeaways

Best clean accuracy: CNN from scratch (69.1%) — small and effective
Best robustness: Noise-augmented CNN — smallest accuracy drop (17.6%) with competitive clean performance
Transfer learning: VGG-16 provides strong results with minimal training, but offers no noise robustness
Most efficient: CNN from scratch — 2.4M params vs 138M+ for VGG

07Model Output Comparison

Clean vs noisy test output across all models.