CNN-Based Image Classification

This repository contains multiple implementations of Convolutional Neural Networks (CNNs) for image classification tasks. The models range from custom-built CNN architectures to transfer learning using pre-trained networks like ResNet50. This repository is structured for flexibility, enabling experimentation with different architectures and datasets.

Features

1. Custom CNN Models:

   • Designed from scratch for specific datasets.
   • Includes feature extraction using multiple convolutional and pooling layers.
   • Augmentation and regularization to prevent overfitting.

2. Small CNN Model:

   • A simplified CNN architecture suitable for small datasets or quick experiments.
   • Focuses on achieving a balance between performance and computational efficiency.

3. Transfer Learning:

   • Uses ResNet50 as the base model for robust feature extraction.
   • Custom dense layers added for classification tasks.
   • Includes fine-tuning for improved performance.

4. Data Augmentation:

   • Rotation, zoom, and flip augmentations implemented to enhance dataset variability.
   • Real-time augmentation during training using ImageDataGenerator.

5. Evaluation:

   • Supports metrics like accuracy, confusion matrices, and loss/accuracy plots.
   • Clear separation of train, validation, and test datasets to prevent data leakage.

6. Callbacks:

   • Early stopping to prevent overfitting.
   • Learning rate reduction on plateau.

---

Prerequisites

• Python 3.8 or higher
• TensorFlow 2.x
• NumPy
• Matplotlib
• scikit-learn

Install the required dependencies with:

pip install -r requirements.txt

---

Directory Structure

.
├── data
│   ├── train
│   ├── validation
│   └── test
├── models
│   ├── custom_cnn.py
│   ├── small_cnn.py
│   ├── transfer_learning_resnet50.py
├── notebooks
│   └── training_visualization.ipynb
├── requirements.txt
├── README.md
└── utils
    ├── data_preprocessing.py
    └── plot_metrics.py

---

Getting Started

1. Prepare the Dataset

Organize the dataset into the following structure:

.
├── train
│   ├── class_1
│   ├── class_2
│   └── ...
├── validation
│   ├── class_1
│   ├── class_2
│   └── ...
└── test
    ├── class_1
    ├── class_2
    └── ...

2. Train a Model

Run any of the scripts in the models/ directory. For example, to train the custom CNN model:

python models/custom_cnn.py

3. Evaluate the Model

Visualize metrics and evaluate model performance using the provided utilities. For example:

python utils/plot_metrics.py

---

Models Overview

Custom CNN Model

A robust architecture with multiple convolutional layers, batch normalization, and dropout for regularization.

Small CNN Model

A lightweight architecture ideal for quick experimentation and smaller datasets.

Transfer Learning with ResNet50

A state-of-the-art approach leveraging pre-trained weights of ResNet50 for feature extraction and fine-tuning.

Results

Include details about the performance of different models, such as:

• Accuracy
• Precision, Recall, F1-Score
• Confusion Matrices

Contributing

Feel free to fork this repository and submit pull requests. Contributions in the form of bug fixes, feature additions, or performance enhancements are welcome.