CrackESS: A Self-Prompting Crack Segmentation System for Edge Devices

摘要

Structural Health Monitoring (SHM) is fundamental to infrastructure maintenance, enabling the early detection of defects and ensuring long-term safety. Crack segmentation serves as a critical technique within SHM for assessing structural health. Recent advancements in deep learning have shown remarkable performance in this area and contributed to the field of automated inspection. However, the diverse characteristics of cracks and complex environmental backgrounds pose significant challenges to accurate and robust crack segmentation. In addition, the high computational demands of most models hinder practical deployment on resourceconstrained edge devices. To address these issues, we propose CrackESS, a novel self-prompting system for accurate and efficient concrete crack segmentation. In this paper, we leverage a YOLOv8n model for prompt generation and introduce a LoRA-based fine-tuning strategy to obtain a lightweight SAM model for crack segmentation. We further propose a Crack Mask Refinement Module (CMRM) to improve segmentation quality. We evaluate CrackESS on three datasets (Khanhha dataset, Crack500, and CrackCR) and validate its practical feasibility through deployment on a climbing robot system for real-world inspection tasks.