Lightweight Segmentation Model for Real-Time Detection of Multi-Type Façade Defects Using Enhanced YOLOv12
TL;DR
A lightweight instance segmentation model for real-time multi-type façade defect detection, built upon an enhanced YOLOv12n architecture that improves efficiency and robustness through parameter sharing, scale adaptation, and group normalization is proposed.
Abstract
Accurate segmentation of façade defects is essential for quantifying damage extent and supporting maintenance decisions in urban infrastructure management. Accurate segmentation of façade defects is essential for extracting visible damage regions and providing auxiliary information for urban infrastructure maintenance. However, existing detection methods based on bounding boxes often fail to delineate defect boundaries precisely, limiting their practical applicability. To address this challenge, this study proposes a lightweight instance segmentation model for real-time multi-type façade defect detection, built upon an enhanced YOLOv12n architecture. Three task-oriented modules are incorporated and adapted for façade defect segmentation: a Global Edge Information Transfer (GEIT) module that enhances boundary representation by integrating multi-scale edge features into shallow layers; a Hierarchical Attention Fusion Block (HAFB) that captures both local details and global semantics via parallel attention and re-parameterizable convolutions; and a Lightweight Shared Convolutional Segmentation Head (LSCSH) that improves efficiency and robustness through parameter sharing, scale adaptation, and group normalization. Experimental results show that the proposed model achieves 91.5% precision, 84.7% recall, 87.2% mAP@0.5, and 64.8% mAP@0.5:0.95, with a compact size of 5.1 MB and real-time speed of 201 FPS, demonstrating strong segmentation performance and deployment potential.