基于改进YOLOv8n的带钢表面缺陷检测方法

Surface defect detection method of strip steel based on improved YOLOv8n

针对带钢表面缺陷种类多样、特征不明显,导致漏检和错检等问题,提出一种改进YOLOv8n的带钢表面缺陷检测方法。首先,为适应较小尺寸目标,增加P2检测层来识别各类缺陷,减少漏检率,以及设计一种高效的PConv检测头,维持推理速度;其次,采取将YOLOv8n颈部中的C2f模块和可变形卷积DCNv2融合的方式,增强模型特征提取能力;此外,在骨干网络输出层引入大动态选择性模块LSKNet,来扩大模型的感受野并提高目标检测的准确性;最后,选择SIoU损失函数替换CIoU损失函数,增强网络收敛效果,从而提高识别精度。改进后YOLOv8n方法在CSU_STEEL数据集上测试,实验结果表明,平均精度均值(mAP)mAP@0.5比原模型提高8.6%,达到82.3%,体积只增加0.5 MB。改进后的方法对带钢表面缺陷有更好检测结果,可为带钢缺陷检测方法的研究提供参考意义。

Aiming at the problems of strip steel surface defects with various types and inconspicuous features,which lead to leakage and misdetection,an improved YOLOv8n strip steel surface defect detection method is proposed.Firstly,to adapt to smaller size targets,a P2 detection layer is added to identify various types of defects and reduce leakage detection,as well as an efficient PConv detection head is designed to maintain the inference speed.Secondly,a fusion of the C2f module in the neck of YOLOv8n and the deformable convolutional DCNv2 is adopted to enhance the feature extraction capability of the model.Furthermore,a large dynamics selective module is introduced in the output layer of the backbone network LSKNet,to expand the sensory field of the model and improve the accuracy of target detection.Finally,the SIoU loss function is chosen to replace the CIoU loss function to enhance the network convergence effect,thus improving the recognition accuracy.The improved YOLOv8n method is tested on the CSU_STEEL dataset,and the experimental results show that the mAP@0.5 is improved by 8.6%the original model to 82.3%,and the volume only increases by 0.5 MB.The improved method has better detection results for strip surface defects,which can provide a reference significance for the research of the defect detection method of strip steel.