mernaabass/Egypt-Vision-AI-Powered-Landmark-Recognition

An AI-powered image classification system for Egyptian landmarks using Transfer Learning and Fine-Tuning with MobileNetV2. Developed for the AI424 - Deep Learning course. # 🏛️ Egypt-Vision: AI-Powered Landmark Recognition **Course Project: AI424 - Deep Learning** **Academic Year:** 2025/2026 **Institution:** Faculty of Artificial Intelligence --- ## 🌟 Overview **Egypt-Vision** is a comprehensive end-to-end training pipeline designed to classify iconic Egyptian landmarks. This project demonstrates the practical application of **Deep Learning** and **Computer Vision** to tourism and cultural heritage, using a lightweight architecture optimized for real-world efficiency. --- ## 📝 Compliance with Task Requirements **Requirement:** > *"Write your own code to modify any pre-trained CNN models and make the suitable fine-tuning to fit your case study."* **Achievement & Implementation:** * **Custom Model Architecture:** I selected **MobileNetV2** as the base model for its efficiency. I modified the architecture by removing the top classification layers (`include_top=False`) and integrating a custom head consisting of: * `GlobalAveragePooling2D`: To reduce spatial dimensions. * `Dropout (0.3)`: To prevent overfitting during the fine-tuning stage. * `Dense (Softmax)`: Tailored specifically to classify the selected Egyptian landmarks. * **Advanced Fine-Tuning Strategy:** I implemented a **Triple-Stage training process** rather than a standard one-step fit: 1. **Stage 1 (Feature Extraction):** Training only the custom head while keeping the base frozen. 2. **Stage 2 (Partial Fine-Tuning):** Unfreezing the top 30 layers for architectural feature adaptation. 3. **Stage 3 (Full Unfreeze):** Unfreezing the entire network with a microscopic learning rate ($5 \times 10^{-6}$) to achieve maximum precision without destroying pre-trained weights. --- ## 🚀 Technical Highlights * **Base Model:** MobileNetV2 (Pre-trained on ImageNet). * **Optimization:** Adam Optimizer with Progressive Learning Rate Decay. * **Regularization:** Early Stopping (Restoring best weights) and Real-time Data Augmentation. * **Performance:** Achieved **70% Validation Accu …