基于改进YOLOv8n的钢表面缺陷检测模型

A Steel Surface Defect Detection Model Based on the Improved YOLOv8n

针对钢表面缺陷尺寸微小且与背景重合度高而导致的传统缺陷检测算法存在漏检和错检的问题,本文提出了一种基于YOLOv8n改进的钢表面缺陷检测算法。首先,使用全维动态卷积ODConv构建C2f-ODConv模块,在多维度提取钢表面的缺陷特征,增强网络全域特征提取的能力,同时引入基于NWD(Normalized Wasserstein Distance)度量方式的损失函数,提高网络对钢表面缺陷的定位准确度;其次,通过减少网络头部卷积次数,提出新的检测头Light-Detect减少模型的资源占有率,提高模型检测的实时性;最后,在SPPF后加入CBAM注意力机制从而在不同图层上提高特征之间的耦合性。实验结果表明,文中所提算法YOLOv8n-Eff在钢表面缺陷数据集NEU-DET上的mAP(mean Average Precision)值达到78.6%,与YOLOv8n算法相比较,mAP提高了3.2%,计算量减少了2.4 G,并且裂纹目标缺陷的AP值提升了10.8%。结果验证了YOLOv8n-Eff算法可以提升钢表面目标缺陷的平均检测精度,降低漏检率并减少模型计算量,有效满足钢表面缺陷检测的需求。

Due to the small size of the surface defects on steel and their high overlap with the back-ground,traditional defect detection algorithms face issues of missing and wrongly detecting these defects.First,the C2f-OConv module was constructed using full-dimensional dynamic convolution(ODConv).This module was designed to extract defect features from multiple dimensions,enhanced the network's global feature extraction capabilities.A loss function based on Normalized Wasserstein Distance(NWD)was introduced.This function aimed to boost the positioning accuracy of the network for steel surface de-fects.Additionally,by reducing the number of convolution operations in the network head,a new detec-tion head called Light-Detect was proposed,it decreased the model’s resource occupancy and improved the model’s real-time detection capabilities.The CBAM attention mechanism was added after SPPF to enhance feature coupling across different layers.Experimental results demonstrated that the mAP(mean average precision)value of the YOLOv8n-Eff algorithm was obtained.It reached 78.6%on the NEU-DET steel surface defect dataset.This represented a 3.2%increase in mAP compared to the original YOLOv8n algorithm.The computational cost was reduced by 2.4 G,and the AP value for crack target defects improved by 10.8%.The results confirmed that the YOLOv8n-Eff algorithm improved average de-tection accuracy for target defects on steel surfaces,reduced the missed detection rate,and lowered the model’s computational costs,effectively meeting the requirements for steel surface defect detection.