Tipo di tesi
Tesi di laurea magistrale
Titolo
Development of Deep Learning Systems for Vehicle Damage Detection and Segmentation
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Parole chiave
- cloud-based deployment
- damage localization
- deep learning
- instance segmentation
- insurance automation
- multi-instance segmentation
- real-time inference
- scalable AI systems
- vehicle damage assessment
Data inizio appello
27/05/2025
Consultabilità
Non consultabile
Data di rilascio
27/05/2095
Riassunto (Italiano)
Automated vehicle damage detection is increasingly critical for enhancing the efficiency of insurance claims and reducing manual inspection efforts. This research presents a YOLOv11-based instance segmentation system tailored to identify and segment various damage types as dents, scratches, cracks, tireflat, lamp broken and glass shatter under real-world conditions including occlusions and varying lighting. To ensure robustness, the model is trained on a diverse, multi-source dataset combining CarDD and internal inspection data from Wondersys Srl. The proposed system achieves real-time, cloud-based inference using Google Vertex AI, enabling scalable deployment. Experimental evaluations reveal that YOLOv11 outperforms models such as Mask R-CNN, U-Net, DeepLabV3, and prior YOLO versions (v8, v9) in both segmentation accuracy and inference speed. The model attains 85% mAP@0.5 and high IoU scores across multiple damage classes, confirming its suitability for practical, high-throughput applications. This work delivers a cost-effective and scalable AI solution for automated damage assessment, offering significant improvements for modern insurance and fleet management systems.
