YOLO-ML:基于多尺度特征层注意力机制的滑轨缺陷检测方法

YOLO-ML:Sliding rail defect detection method based on multi-scale feature layer attention mechanism

汽车拉索滑轨外观质量是保障车窗升降系统良好工作的重要保障,外观缺陷会导致车窗升降系统晃动、异响、卡顿、腐蚀等问题,因此对汽车拉索滑轨外观进行缺陷检测至关重要。提出了一种用于汽车拉索滑轨外观缺陷的检测方法,该方法使用YOLO V8作为基础模型,通过加入小目标检测头、多尺度序列特征金字塔(multi-scale sequence feature based feature pyramid network,multi-SSFPN)模块、多尺度特征层注意力(layer-attention)模块,实现了滑轨外观多尺寸缺陷的实时检测。实验和现场测试结果表明,该方法在滑轨缺陷数据集上的平均检测精度Box-mAP50为91.5%,Mask-mAP50为89.4%,能够很好地识别滑轨外观缺陷。为了进一步验证了该方法的有效性,在公开数据集NUE-DET和GC 10-DET上对方法进行了评估,分别取得了平均检测精度78.5%和70.8%,与当前数据集上最好效果相比取得了有竞争力的结果。

The appearance quality of the car window regulator rail is an important guarantee for the proper functioning of the window lifting system.Appearance defects can lead to problems such as shaking,abnormal noise,jamming,and corrosion in the window lifting system.Therefore,it is crucial to detect defects in the appearance of car window regulator rails.This paper proposes a method for detecting defects in the appearance of car window regulator rails.The method uses YOLO V8-p2 as the base model and incorporates a small target detection head,multi-SSFPN feature pyramid network,and multi-scale feature layer attention module to achieve real-time detection of defects of various sizes in the rails’appearance.Experimental and on-site test results demonstrate that the proposed method achieves an average detection precision of Box-mAP50 at 91.5%and Mask-mAP50 at 89.4%on the rail defect dataset,effectively identifying defects in the appearance of the rails.To further validate the effectiveness of the proposed method,evaluations are conducted on publicly available datasets NUE-DET and GC 10-DET,yielding average detection precisions of 78.5%and 70.8%,respectively.These results demonstrate competitive performance compared to the best results on the current dataset.