🧠 Image Classification from Scratch using CNN

This deep learning project demonstrates the end-to-end development of an image classification system using a custom Convolutional Neural Network (CNN) architecture, built entirely from scratch — without relying on pre-trained models.

📌 Key Achievements:
✅ Built CNN model from zero (no transfer learning)
✅ Achieved 96% test accuracy on unseen data
✅ Designed a robust pipeline including cleaning, augmentation, training, evaluation, and deployment
✅ Deployed on Hugging Face with live API access

---

📁 Project Contents

Notebook	Description
01_data_analysis.ipynb	Initial EDA, class distribution analysis, imbalance insights
02_data_augmentation_static.ipynb	Data augmentation techniques to address class imbalance
03_model_attempt1_88acc.ipynb	First baseline CNN model with 88% accuracy
04_model_final_96acc.ipynb	Final refined CNN model with 96% accuracy
class_info.json	Contains general descriptive information about each class, used to display names during testing and inference
deployment/	Scripts and links for online inference API
example_request.ipynb	Upload an image and send it to the deployed Hugging Face API for testing

🗒️ Note: The notebooks 01_data_analysis.ipynb and 02_data_augmentation_static.ipynb contain some printed messages in Arabic. This does not affect any code functionality or the results.

---

🧼 Dataset Preparation & Cleaning

The dataset was originally based on publicly available resources, including:

• Egypt Monuments Dataset on GitHub

Additionally, more images were manually collected from various online websites to enrich and balance the dataset.

• Broken/corrupted images were detected and removed.
• Validated file formats and color channels (e.g., RGB).
• Watermarked images were removed manually.
• Clean replacement images were manually merged into the dataset post-cleaning.

🧾 Note: If the merge step is not visible in the code, it's because it was done offline to ensure full control over dataset quality before training.

---

• After the manual collection and cleaning steps, data augmentation techniques were applied locally on the dataset to synthetically increase the number of images per class. This helped to balance the classes and improve model generalization during training.

---

🖼️ Dataset Versions

There are two versions available for download:

Version	Description	Download
🧹 Raw Cleaned Dataset	Contains original cleaned images without augmentation or splitting. Useful if you'd like to redo preprocessing and augmentation from scratch.	Download from Google Drive
📦 Final Dataset	Augmented dataset with ~1000 images per class, already split into train, val, and test. Used to train the final model.	Download from Kaggle

⚠️ These datasets are used for educational purposes only and were collected from publicly available sources.

---

🧠 CNN Model Architecture Overview

The deep learning model is a custom-built Convolutional Neural Network (CNN) designed from scratch, without using any pre-trained weights.

Architecture Summary:

• The network starts with a convolutional layer of 32 filters, followed by progressively deeper convolutional layers with 64, 128, and 256 filters.
• After some of the convolutional layers, MaxPooling is applied to reduce spatial dimensions.
• The feature maps are then flattened and passed through a fully connected dense layer with 512 neurons, followed by a dropout layer (rate 0.1) to reduce overfitting.
• The final output layer uses a softmax activation with 21 units, corresponding to the 21 classes in the dataset.

Training Details:

• Two models were trained during experimentation:
  • An initial model achieving 88% accuracy.
  • The final model, which reached 96% accuracy on the test set.
• Early stopping was applied manually at epoch 52 out of a planned 75 epochs, based on monitoring validation performance.

For detailed implementation and code, please refer to the notebook:

• 04_model_final_96acc

---

📈 Performance

Metric	Value
Test Accuracy	96% on unseen data
Loss	Low & stable
Overfitting	Avoided using augmentation & dropout
Epochs Trained	52 (out of 75 planned epochs)
Early Stopping	Manual – Training was stopped at epoch 52 after observing a plateau in validation loss and no further improvement in accuracy

📌 Manual early stopping was used by closely monitoring training and validation performance. Although 75 epochs were planned, training was stopped at epoch 52 when the model stabilized and began to show signs of potential overfitting.

---

🏷️ Supported Classes

The model is trained to classify exactly 21 distinct ancient Egyptian landmarks, including royal statues, temples, and pyramids.

Here are the supported classes:

• King Akhenaten
• King Amenhotep III
• Bent pyramid of Senefru
• Colossi of Memnon
• Goddess Isis
• Queen Hatshepsut
• Khafre Pyramid
• King Thutmose III
• King Tutankhamun
• Queen Nefertiti
• Pyramid of Djoser
• King Ramesses II
• Ramessum (Memorial Temple of Ramesses II)
• King Zoser
• Tutankhamun with Ankhesenamun
• Temple of Hatshepsut
• Temple of Isis in Philae
• Temple of Kom Ombo
• The Great Temple of Ramesses II
• Menkaure Pyramid
• Sphinx

⚠️ The model cannot predict labels outside of these 21 predefined classes.

---

🚀 Deployment

The final trained model has been deployed and made publicly available for real-time inference.

• 🤖 Model on Hugging Face:
  🔗 View on Hugging Face

• 🧠 Live Inference API:
  🔗 Access the API

⚠️ Note for Users:
To interact with the live inference API directly, please use a tool like Postman or a script (e.g., using Python's requests library).
Web browsers cannot send file uploads in POST requests correctly, so using Postman is necessary to test the /predict endpoint with image files.

🚀 Try it on Google Colab

⚠️ Note: Only .jpg, .jpeg, .png, .bmp and .webp image formats are supported.
Please convert other formats (such as .heic) before uploading to avoid errors.

🖼️ Sample Image for Testing

To help you try the live API quickly, here's a sample image you can use:

King_Zoser

📎 You can right-click the image to save it and use it in the Google Colab demo or Postman.