YOLOv12-VSD: A Transfer-Learning-Assisted Real-Time Detection Algorithm for Vehicle Surface Defects

Vehicle surface defect detection faces three core challenges: classification-localization inconsistency for boundary-sensitive defects, insufficient multi-scale feature response across defect sizes, and cross-scenario generalization degradation caused by domain shift among production lines. This paper proposes YOLOv12-VSD, an improved detection algorithm addressing these issues through coordinated modifications at three levels. An IoU-aware classification loss aligns classification confidence with localization quality. A reparameterized convolution module at the P4 feature level (P4-RepC3) enriches intermediate-layer directional feature diversity without increasing inference cost. A multi-scale spatial pyramid pooling-fast structure at the P5 feature level (P5-SPPF) expands the effective receptive field for large-area defects. A three-stage transfer learning framework comprising source-domain pretraining, target-domain adaptation, and low-learning-rate refinement is further designed to reduce domain shift with limited annotations. Experiments show that YOLOv12-VSD achieves a mean Average Precision at IoU threshold 0.50 (mAP@50) of 0.715, the highest among six comparison models, with only 6.1M parameters and 17.1 giga floating-point operations per second (GFLOPs). After three-stage transfer, mAP@50 improves from 0.531 to 0.652, with training duration reduced by 64%.