I need an automated way to inspect every category of EU traffic signs from cropped images and decide two things for each crop: 1) whether the sign shows fading/discoloration or graffiti/vandalism, and 2) how severe that damage is on a clear, explainable scale. You may start from an existing classifier or build one from scratch, as long as the final solution processes batches of images and outputs a CSV (image ID, detected damage type, damage grade). I will supply a small, manually-labeled sample set; you will expand on it with augmentation or transfer learning to reach reliable generalisation across all regulatory, warning, and guide signs used in the EU. Key expectations • A documented training pipeline (Python + PyTorch/TensorFlow or similar) • A reproducible inference script or notebook that runs on my machine (Ubuntu + GPU) • A short report summarising accuracy metrics and confusion areas, plus suggestions for future improvement When you reply, show me one previous computer-vision project—preferably anything to do with classification or defect detection—so I can judge fit quickly.