Aircraft Damage Classification & Captioning

Overview

This project automates aircraft damage assessment from images using deep learning. It first classifies damage type (dent vs crack) using a VGG16 pre-trained model, then generates descriptive captions and summaries using the BLIP transformer model. This workflow simulates intelligent maintenance support systems used in aviation safety and MRO (Maintenance, Repair, Overhaul) operations.

Key Features

• Classifies aircraft surface damage as dent or crack
• Uses VGG16 for feature extraction and transfer learning
• Generates human-readable damage captions and summaries via BLIP
• Includes training curves, evaluation metrics, and visual predictions
• Reproducible pipeline with dataset loading, preprocessing, and labeling

Tech Stack

• TensorFlow / Keras
• VGG16 Pre-trained CNN
• BLIP Transformer (Hugging Face)
• ImageDataGenerator
• Matplotlib / NumPy

Model Performance

• Achieved ~81% accuracy on test set (5 epochs, frozen VGG16 base)
• Accurate visual label predictions on test images
• BLIP generates contextual captions and damage summaries

Applications

• Automated inspection support
• Faster maintenance documentation
• Visual defect triaging systems
• Aviation safety research and prototypes

Next Steps

• Fine-tune VGG16 layers for higher accuracy
• Expand dataset (different aircraft types, lighting, angles)
• Deploy as a web dashboard with file upload
• Integrate OCR for serial number tagging

Acknowledgements

• VGG16: ImageNet pre-trained model
• BLIP: Salesforce / Hugging Face
• Aircraft dataset: CC BY 4.0