Lightweight BGS-YOLOv8 for surface crack detection in civil engineering infrastructure

To address the problems of false detection and excessive parameters in surface crack detection for civil infrastructure inspection, we propose an improved BGS-YOLOv8 method. First, we incorporate singular value decomposition based on cumulative contribution rate into the C2f modules to reduce model parameters while preserving important crack features. Second, we design a multi-dimensional channel importance evaluation model and an adaptive backtracking recovery method to compensate for the accuracy loss caused by over-compression. Experimental results on the RDD2022 dataset demonstrate that BGS-YOLOv8 improves mAP@0.5 by 2.33% over YOLOv8, while reducing parameters by 22.92% and GFLOPs by 18.29%. Although the current validation is conducted on road-surface images, the lightweight design is intended to be applicable to other surface crack inspection tasks, such as concrete structures.