基于GAB-YOLO的钢材表面缺陷检测方法

Steel Surface Defect Detection Method Based on GAB-YOLO

摘要:

针对钢材表面缺陷识别中因形状复杂、尺度差异大及对比度低导致的精度瓶颈、漏检误检及计算开销大等问题；对YOLO11n进行了结构优化与创新,构建了GAB-YOLO检测模型；设计了集成迭代注意力特征融合机制与分组空间卷积的C3K2-GSIA模块,实现对关键缺陷特征的关注增强与计算复杂度的降低；提出采用双分支架构的自适应分离下采样模块ASDown,完成下采样参数的自适应调整与微小缺陷细节特征的保留；建立融合多分支辅助特征金字塔与加权双向金字塔优点的BiMAFPN网络,利用加权融合机制提升小目标特征的表达权重；采用PIoUv2损失函数替代原有CIoU损失,实现锚框回归效率的提升与关键样本的聚焦；经实验测试,在mAP@0.5指标上达到81.5%,精确率与召回率分别达到77.8%与74.2%；相较于YOLO11n基准模型,mAP@0.5提升6.7个百分点,参数量与计算量分别降低11.24%和7.81%；由此可见,该模型满足对钢材表面缺陷检测的应用需求。

Abstract:

Research was conducted on the precision bottlenecks, missed detections, false positives, and high computational costs in steel surface defect recognition caused by complex shapes, significant scale variations, and low contrast; structural optimization and innovation were performed on YOLO11n architecture to construct GAB-YOLO detection model; a C3K2-GSIA module integrating the iterative attention feature fusion (iAFF) mechanism and grouped spatial convolution (GSConv) was designed to enhance focus on key defect features and reduce computational complexity; an adaptive separation downsampling (ASDown) module utilizing a dual-branch architecture was proposed to achieve adaptive adjustment of downsampling parameters and preservation of detailed features for tiny defects; a bidirectional multi-branch auxiliary feature pyramid network (BiMAFPN) fusing the advantages of the multi-branch auxiliary feature pyramid network (MAFPN) and the bidirectional feature pyramid (BiFPN) was established to increase the representation weights of small target features through a weighted fusion mechanism; the Powerful-IoU version 2 (PIoUv2) loss function was adopted to replace the original complete intersection over union (CIoU) loss to improve anchor box regression efficiency and focus on key samples; experimental testing achieved a mean average precision (mAP) at a 0.5 threshold (mAP@0.5) of 81.5%, with precision and recall reaching 77.8% and 74.2%, respectively; compared with the YOLO11n baseline model, the mAP@0.5 was improved by 6.7 percentage points, while the parameter count and computational load were reduced by 11.24% and 7.81%, respectively; the results demonstrate that the model satisfies the application requirements for steel surface defect detection.