📌 Project Description
This project implements a Computer Vision system based on Deep Learning designed for the automatic classification of date quality. The primary focus is identifying "Loose Skin" (skin detachment), a common quality defect, distinguishing deteriorated fruits from healthy ones.
The application integrates a neural model trained in a cloud environment (Google Colab) with an interactive web interface developed in Streamlit for local inference.
🧠 Model Architecture
To maximize accuracy with an optimized dataset, Transfer Learning was used. Starting from a network pre-trained on ImageNet and adding specific final layers to adapt the model to the available dataset.
🛠️ Technical Challenges
Cross-Version Compatibility
Resolved incompatibility between Keras 3 (Colab) and Keras 2 (local environment) by loading weights-only via layer name mapping.
User Experience (UX)
Enhanced the Streamlit interface with custom CSS for a clean design, centered buttons, and clear visual feedback.
📈 Future Developments
- Fine-tuning: Unlocking late blocks of DenseNet to specialize the model on specific date textures.
- Data Augmentation: Increasing dataset variety to handle different lighting conditions.
Try the Live Demo
Launch App
🔒 Code Note: For academic compliance, the source code and instances are private. The application demo remains publicly available.