A lightweight multi-scale selective detector for UAV-based building facade defect detection

Accurate detection of building facade defects is essential for urban safety and maintenance decision-making, but UAV-based inspection requires models that remain accurate under limited onboard computing resources. This study proposes EMS-YOLO, a lightweight multi-scale selective detector reconstructed from the YOLOv8 baseline for facade defect detection. The model integrates Partial Convolution (PConv) to reduce redundant spatial computation, a Large Kernel Selective attention block (LSKBlock) to enhance adaptive multi-scale feature representation, and a Content-Aware ReAssembly of FEatures (CARAFE) module to improve the boundary perception of small and blurred defects. A self-built dataset was constructed from UAV and ground close-range images, with original image/building-site-level splitting before augmentation to avoid data leakage. Public datasets, UAV video streams, and embedded platforms were further used for generalization and deployment evaluation. Experimental results show that EMS-YOLO achieves 76.4% mAP@0.5 with 10.1 M parameters and 105.6 FPS on an RTX 3070 Laptop GPU. The proposed method improves crack, spalling, and delamination detection and provides a practical visual screening tool for UAV-based facade inspection.