CIFAR-10 Image Classification with CNN and ResNet50 Transfer Learning

Project Overview

This project explores image classification on the CIFAR-10 dataset using two different deep learning approaches:

• A custom Convolutional Neural Network (CNN) trained from scratch.
• A transfer learning approach based on a pretrained ResNet50 model.

The objective is to classify RGB images into one of ten mutually exclusive categories and evaluate the impact of transfer learning on low-resolution image classification.

The project includes:

• Exploratory Data Analysis (EDA)
• Data preprocessing
• Baseline CNN modeling
• Transfer learning with ResNet50
• Data augmentation
• Selective fine-tuning
• Error analysis and model evaluation

---

Final Results

Model	Test Accuracy
Baseline CNN	60.5%
ResNet50 Head Only	88.7%
ResNet50 Fine-Tuned	91%

Transfer learning improved test accuracy from 60.5% to 91%.

---

Dataset

The project uses the CIFAR-10 dataset available through TensorFlow/Keras.

Dataset characteristics:

• 60,000 RGB images
• Image size: 32 × 32 pixels
• 10 classes
• 50,000 training images
• 10,000 test images

Classes:

• Airplane
• Automobile
• Bird
• Cat
• Deer
• Dog
• Frog
• Horse
• Ship
• Truck

To reduce training time on local hardware, the training dataset was limited to 10,000 images.

---

Dataset Samples

Visual inspection revealed:

• High intra-class variability
• Different viewpoints and object scales
• Complex and inconsistent backgrounds
• Limited visual detail due to the 32×32 resolution

These characteristics make CIFAR-10 a challenging image classification benchmark.

---

Methodology

1. Exploratory Data Analysis

The EDA phase focused on:

• Class distribution
• Sample visualization
• RGB channel analysis
• Pixel intensity distributions

The dataset was found to be well balanced across all classes.

---

2. Baseline CNN

A custom CNN was implemented to establish a performance baseline.

Architecture:

• Conv2D (32) + MaxPooling
• Conv2D (64) + MaxPooling
• Conv2D (128) + MaxPooling
• Dense (128)
• Dropout
• Softmax Output Layer

Total trainable parameters:

~357,000

The CNN was trained directly on the original 32×32 images.

---

3. Transfer Learning with ResNet50

A pretrained ResNet50 model was used as the feature extraction backbone.

Pipeline:

Input Image
     ↓
Resize to 224×224
     ↓
ResNet50 Preprocessing
     ↓
Pretrained ResNet50 Backbone
     ↓
Custom Classification Head
     ↓
Prediction

Custom classification head:

• GlobalAveragePooling2D
• Dense (256)
• Dropout (0.5)
• Softmax Output Layer

Total model parameters:

~24 million

---

4. Training Strategy

The transfer learning workflow consisted of two stages.

Stage 1 — Head-Only Training

• ResNet50 backbone frozen
• Only the custom classification head trained

Result:

88.7% test accuracy

Stage 2 — Selective Fine-Tuning

• Early ResNet50 layers remained frozen
• Deeper layers were unfrozen
• Smaller learning rate applied

Additional techniques:

• Data augmentation
• Dropout regularization
• ReduceLROnPlateau callback

Result:

91% test accuracy

---

Data Augmentation

To improve generalization and reduce overfitting, several augmentation techniques were applied:

• Random horizontal flip
• Random crop / zoom
• Brightness variation
• Contrast variation

These transformations increase visual diversity without changing image labels.

---

Fine-Tuning Performance

The training curves show stable convergence during fine-tuning and good validation performance throughout training.

---

Error Analysis

Confusion Matrix

The confusion matrix shows strong performance across most classes.

Most errors occur between visually similar categories, including:

• Cat ↔ Dog
• Automobile ↔ Truck
• Horse ↔ Deer

These mistakes are expected due to:

• Low image resolution
• Similar textures and shapes
• Ambiguous viewpoints
• Missing contextual information

---

Key Findings

• Transfer learning significantly outperformed the baseline CNN.
• Pretrained ImageNet features transferred effectively to CIFAR-10.
• Data augmentation improved generalization.
• Selective fine-tuning provided additional performance gains.
• Error analysis revealed meaningful semantic confusions rather than random failures.

---

Project Structure

cifar10-image-classification/
│
├── notebooks/
│   ├── 01_gathering_eda.ipynb
│   └── 02_preprocessing_modeling_evaluation.ipynb
│
├── images/
│   ├── dataset_samples.png
│   ├── augmentation.png
│   ├── resnet_finetuning.png
│   └── confusion_matrix.png
│
├── presentation/
│   └── computer_vision_image_classification.pdf
│
├── README.md
├── requirements.txt
└── .gitignore

---

Technologies Used

• Python
• TensorFlow / Keras
• NumPy
• Pandas
• Matplotlib
• Seaborn
• Scikit-Learn

---

Installation

Clone the repository:

git clone https://github.com/your-username/cifar10-image-classification.git
cd cifar10-image-classification

Install dependencies:

pip install -r requirements.txt

---

Running the Project

Open the notebooks in the following order:

1. 01_gathering_eda.ipynb
2. 02_preprocessing_modeling.ipynb

---

Limitations

• Only 10,000 training images were used.
• Training was limited by local hardware resources.
• ResNet50 is relatively large for CIFAR-10.

---

Future Improvements

• Train on the full CIFAR-10 dataset
• Experiment with EfficientNet architectures
• Explore Vision Transformers (ViTs)
• Apply stronger augmentation strategies
• Perform hyperparameter optimization

---

Conclusion

This project demonstrates how transfer learning can dramatically improve image classification performance on low-resolution datasets.

Despite using only a 10,000-image training subset, the fine-tuned ResNet50 achieved more than 91% test accuracy, significantly outperforming a CNN trained from scratch.

The results highlight the effectiveness of pretrained visual representations, data augmentation, and selective fine-tuning in modern computer vision workflows.