Lightweight Detection Network for Small Target Defects in Shaft Lining Crack

Stress‐induced cracks in shaft linings, if left unmonitored during their developmental stages, may ultimately lead to structural rupture and severe safety incidents. To address this critical issue, this study proposes S‐YOLOv8n, a lightweight intelligent detection model designed for the automated and precise identification of micro‐cracks in shaft linings. The proposed framework incorporates two key innovations: (1) A lightweight convolutional downsampling enhancement (CDE) module that synergistically combines standard convolution, depthwise separable convolution, and ECA attention mechanisms. This architecture effectively mitigates complex background interference while preserving crack features during downsampling, enhancing shallow network feature extraction capabilities. (2) A GS‐C2f feature fusion module that optimizes both spatial characteristics of crack morphology and semantic representations of crack severity, concurrently improving inference efficiency. Experimental validation on a custom shaft lining dataset demonstrates that S‐YOLOv8n achieves 1.4% higher detection accuracy than baseline YOLOv8n while reducing model parameters and computational costs by 32.0% and 27.8%, respectively. These advancements significantly contribute to practical implementations of automated crack detection in underground infrastructure maintenance.