An Economical General Purpose AI & UAV Based Inspection System

Abstract

Manual structural inspection of buildings and civil infrastructure is slow, costly, and hazardous, since it depends on scaffolding, rope access, or elevated work platforms that place technicians at height for extended periods. This paper presents an economical UAV-based inspection system that combines onboard defect detection with cloud-side image restoration, removing both the access equipment and the manual image-review stage from the inspection workflow. A YOLOv11 Nano detector executes directly on a Raspberry Pi 4 carried by the airframe, discriminating flexural cracks, shear cracks, combined damage, and undamaged surfaces at 14 frames per second (approximately 71 ms per frame) with no GPU. Frames containing candidate defects are transmitted to a cloud stage where DeblurGAN removes motion blur and SRGAN recovers fine texture before severity is classified and reported to an IoT dashboard. Trained on nine public structural-damage datasets, YOLOv11n attains mAP@0.5 of 0.538, precision of 0.540, recall of 0.536, and an F1-score of 0.538, compared with 0.627 mAP@0.5 for the larger YOLOv8s and 0.553 for YOLOv8n. YOLOv11n was selected for deployment because it retains approximately 97% of the mAP of YOLOv8n while using around 19% fewer parameters and 25% fewer floating-point operations, and sustains realtime throughput on a hardware platform costing approximately INR 32,000 (USD 377). The restoration stage raises image quality from 17.01 dB PSNR and 0.6415 SSIM after deblurring to 23.80 dB PSNR and 0.7369 SSIM after super-resolution. The results indicate that automated structural screening is feasible on commodity edge hardware in low-resource settings, at a platform cost below that of a single conventional scaffold-based survey.