ResNet & ResNeXt Image Classification Project

This repository contains three Python scripts for training, evaluating, and fine-tuning deep learning models based on ResNet-50 and ResNeXt-50 architectures using TensorFlow/Keras. The goal of the project is to classify rock images into five different geological classes.

---

Repository Structure

├── pre_ResNet-50.py
├── pre_ResNeXt50.py
├── retrain_resnet_resnext.py
├── README.md
└── LICENSE

---

File Descriptions

1. pre_ResNet-50.py

This script performs initial training (transfer learning) using a pre-trained ResNet-50 model.

Key features:

• Uses ImageNet pre-trained weights
• Image data augmentation (shear, zoom, horizontal flip)
• Custom fully connected layers
• SGD optimizer
• Accuracy and loss visualization
• Confusion matrix computation

Classification classes:

• intact Rock
• stylolite
• horizontal plug
• vertical plug
• Crack

---

2. pre_ResNeXt50.py

This script is similar to the ResNet version but uses ResNeXt-50 as the backbone network.

Main differences from ResNet-50:

• Higher cardinality architecture
• Improved feature representation capability

Other steps include:

• Data augmentation
• Training with SGD optimizer
• Model evaluation and confusion matrix visualization

---

3. retrain_resnet_resnext.py

This script is designed for fine-tuning previously trained ResNet or ResNeXt models.

Capabilities:

• Dynamic learning rate scheduling (LearningRateScheduler)
• SGD with momentum
• Dropout layers to reduce overfitting
• Multiple fine-tuning strategies (commented and optional)

Fine-tuning options included:

• Freezing selected base model layers
• Switching optimizers (e.g., Adam)
• Advanced data augmentation
• Adding additional dense layers

---

Requirements

• Python 3.8+
• TensorFlow 2.x
• NumPy
• Matplotlib
• Seaborn
• scikit-learn
• Google Colab (recommended)

Install dependencies:

pip install tensorflow numpy matplotlib seaborn scikit-learn

---

Dataset Structure

The dataset must be organized as follows (Google Drive):

DataSplit/
├── train/
│   ├── intact Rock/
│   ├── stylolite/
│   ├── horizontal plug/
│   ├── vertical plug/
│   └── Crack/
└── validation/
    ├── intact Rock/
    ├── stylolite/
    ├── horizontal plug/
    ├── vertical plug/
    └── Crack/

---

How to Run

1. Open the desired script in Google Colab
2. Mount Google Drive
3. Verify dataset paths
4. Run the script

---

Outputs

• Training and validation accuracy plots
• Training and validation loss plots
• Final validation accuracy
• Confusion matrix for model performance analysis

---

📜 License

This project is released under the MIT License.
You may freely use and adapt the code for research and educational purposes.

See the LICENSE file for full details.