改进YOLOv5的钢材表面缺陷检测网络轻量化研究

Research on Lightweight of Steel Surface Defect Detection Network Based on Improved YOLOv5

在YOLOv5模型的基础上设计了一种改进的轻量化网络,能够快速准确地实现钢材表面缺陷检测。首先,使用基于梯度路径设计的ELAN结构,通过提高网络的学习能力来提高检测精度;其次,引入深度可分离卷积和Ghostv2模块减少模型体积和参数量;最后,利用SIOU边界框损失函数训练模型,使模型能够快速收敛并且精确回归。在NEU-DET上的实验结果表明,改进后的模型mAP值提升到77.0%,相较于原模型提高了5.3%,模型体积减少了42.1%,参数量减少了43.4%,检测速度也快了0.4 ms,实现了模型轻量化效果和检测精度的平衡,为后续在硬件终端上部署提供了一种可行方案。

This research designs an improved lightweight network based on YOLOv5 model,which can quickly and accurately detect steel surface defects.Firstly,the ELAN structure based on gradient path design is used to improve the detection accuracy by improving the learning ability of the network;Secondly,the depth separable convolution and Ghostv2 module are introduced to reduce the volume and parameters of the model;Finally,the SIOU boundary box loss function is used to train the model,so that the model can quickly converge and accurately regress.The experimental results on NEU-DET show that the mAP value of the improved model is increased to 77.0%,which is 5.3%higher than the original model,the model volume is reduced by 42.1%,the number of parameters is reduced by 43.4%,and the detection speed is also 0.4 ms faster,realizing the balance between the lightweight effect of the model and the detection accuracy,and providing a feasible scheme for subsequent deployment on the hardware terminal.