基于HPDE-YOLO的钢材表面缺陷检测算法

Steel Surface Defect Detection Algorithm Based on HPDE-YOLO

针对钢材表面缺陷检测算法精度低、计算量大等问题,提出一种基于YOLOv8n的检测算法HPDE-YOLO(high-level path aggregation dynamic efficient network-YOLO)。首先,引入高效多尺度注意力(efficient multi-scale attention,EMA)机制,与主干部分的C2f融合,增强特征提取能力,并采用C2f-Faster结构提高模型计算速度;其次,提出一种高级筛选双向特征融合金字塔与路径聚合网络(high-level screening-feature bidirectional fusion pyramid and path aggregation network,HS-FPAN),在多个尺度上同时增强语义特征,有效提升模型对细节的捕捉能力;最后,融合动态上采样模块DySample,进一步提升模型检测速度。在NEU-DET数据集上的实验结果表明,相较YOLOv8n模型,HPDE-YOLO模型检测的平均精度均值mAP@0.5达到84.2%,提升了5.7个百分点,裂纹类缺陷检测的平均精度均值mAP提升了26.88个百分点,参数量减少了45%,浮点运算量减少了32%。HPDE-YOLO模型在满足轻量化的同时能够有效提升钢材表面缺陷检测的精度,且易于移动端部署,满足工业生产需求。

To address the issues of low accuracy and high computational demand in steel surface defect detection algorithms,a detection algorithm based on YOLOv8n,named HPDE-YOLO,is proposed.Initially,an efficient multi-scale attention(EMA)mechanism is introduced,which is integrated with the backbone’s C2f to enhance feature extraction capabilities.It is coupled with a C2f-Faster structure to improve the model’s computational speed.Subsequently,a high-level screeningfeature bidirectional fusion pyramid with path aggregation network is proposed,which simultaneously strengthens semantic features at multiple scales,effectively enhancing the model’s ability to capture details.Finally,the model’s detection speed is further increased by incorporating a dynamic upsampling module,DySample.Experimental results on the NEU-DET dataset show that compared to the YOLOv8n model,the HPDE-YOLO model achieves an average mean precision(mAP)of 84.2%at an IoU threshold of 0.5,representing an improvement of 5.7 percentage points.Specifically,the mAP for crack defect detection has increased by 26.88 percentage points.Additionally,the model reduces the number of parameters by 45%and floating-point operations by 32%.The HPDE-YOLO model not only effectively improves the accuracy of steel surface defect detection while being lightweight but also facilitates deployment on mobile devices,meeting the demands of industrial production.