摘要
Accurate valve nozzle detection is an important component of industrial visual inspection systems;however,structural complexity,scale variation,illumination fluctuation,and partial occlusion remain challenging factors that affect detection stability.This study presents YOLOv12-Enhanced,a refined singlestage detection framework developed for industrial valve nozzle scenarios.The proposed approach incorporates three architectural modifications:a RepViT backbone to enhance hierarchical feature representation through structural re-parameterization and global–local modeling,a Spatial Pyramid Pooling Fast(SPPF)module combined with C2PSA attention to strengthen multi-scale contextual feature extraction,and a Global Edge Information Fusion(GEIF)module to integrate shallow edge information with deep semantic features for improved boundary alignment.Experimental evaluation on the Pascal Visual Object Classes(VOC)dataset shows that the proposed model achieves 71.0%mAP50 and 54.4%mAP50–95 under identical training conditions,exceeding the baseline YOLOv12n.Ablation experiments further demonstrate that each module contributes incremental performance gains.Evaluation on a self-constructed valve nozzle dataset consisting of 500 real industrial images indicates stable detection behavior under varying illumination and partial occlusion conditions.The experimental findings suggest that the proposed structural refinements provide a balanced enhancement in feature representation and localization precision while maintaining comparable computational complexity.
