Lightweight YOLOv8-PSN for Real-Time Detection of Underground Structure Defects: A Comparison of Three Importance Criterion Pruning Methods

To address the limited computational capacity of edge devices in underground defect detection tasks, this study utilizes an improved lightweight model, YOLOv8-PSN, as the baseline to systematically compare the compression mechanisms and performance of three structured pruning criteria: DepGraph, LAMP, and Taylor. The three pruning criteria elicited distinct network-restructuring behaviors. DepGraph exhibited a tendency to substantially compress intermediate layers, while LAMP induced more uniformly distributed channel pruning and was correlated with greater loss of high-frequency texture information. In contrast, Taylor pruning displayed an “inverted-pyramid” restructuring pattern. This pattern effectively reduces redundancy in the detection head while preserving the backbone’s robust feature representation capability. Based on comprehensive quantitative metrics and visual analyses, the Taylor-based method achieves significant acceleration of inference under high compression ratios. Crucially, it maintains higher detection confidence and bounding box continuity, demonstrating greater visual robustness to complex interference—such as strong illumination and low-light conditions—particularly when detecting fine cracks. This “metric-visual” discrepancy effectively highlights the limitations of relying solely on mAP for model evaluation in practical engineering. Furthermore, the proposed GPU-oriented channel alignment strategy improved inference efficiency on the tested NVIDIA RTX 5080 workstation GPU; this result is specific to the tested GPU architecture and may differ on embedded platforms. This study provides valuable guidance on the lightweight design and real-time deployment of defect detection models for underground engineering applications.