改进YOLOv8的轻量化钢材表面缺陷检测

PDF下载HTML

Title:

Improvement of YOLOv8 for lightweight steel surface defect detection

作者:

张冀^1,2, 王定邦¹, 曹锦纲^1,2, 杨立然^1,2

1. 华北电力大学 计算机系, 河北 保定 071003;
2. 华北电力大学 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003

Author(s):

ZHANG Ji^1,2, WANG Dingbang¹, CAO Jingang^1,2, YANG Liran^1,2

1. Department of Computer Science, North China Electric Power University, Baoding 071003, China;
2. Engineering Research Center of intelligent Computing for Complex Energy Systems Ministry of Education, North China Electric Power University, Baoding 071003, China

关键词:

轻量级网络; 缺陷检测; YOLOv8n; 多尺度特征融合; 特征提取; 空间金字塔; 目标检测; WIoU

Keywords:

lightweight network; defect detection; YOLOv8n; multiscale feature fusion; feature extraction; spatial pyramid; object detection; WIoU

分类号:

TP391.4；TG115

DOI:

10.11992/tis.202504018

摘要:

针对钢板表面缺陷检测目前存在的严重漏检误检以及边缘设备难部署问题，提出了一种基于YOLOv8的轻量化的钢材表面缺陷检测算法。1)设计了部分卷积门控线性单元(partial convolutional gated linear unit, PGBN)模块来替换BottleNeck模块，只对部分通道进行卷积操作以减少模型的参数；2)使用不同膨胀率的卷积改进空间金字塔池化模块，加强对细粒度特征的提取，并结合全局到局部空间聚合(global-to-local spatial aggregation, GLSA)模块改进BiFPN(bidirectional feature pyramid network)结构，提出了多尺度特征融合网络GLFPN(global-to-local spatial aggregation bidirectional feature pyramid network)，以保留小目标特征，提升模型精度；3)设计了轻量化的检测头，使用共享权重的卷积进一步减少模型的参数量和计算量；4)用WIoU(weighted intersection over union)损失函数来替代原有的损失函数。实验结果表明，改进模型在NEU-DET数据集上mAP50达到了 79.6%，相比YOLOv8n 提升了4.2百分点，而模型的参数量和计算量仅有1.43×10⁶ 和4.7×10⁹，较YOLOv8n分别下降了53.3%和41.9%，在提升准确率的同时更加适合在边缘设备部署。

Abstract:

Surface defects of steel plates still face significant issues of missed and false detections, as well as difficulties in deployment on edge devices. To address these problems, a lightweight steel surface defect detection algorithm based on YOLOv8 is proposed. First, a Partial Convolution Gated Linear Unit (PGBN) module is designed to replace the BottleNeck module, performing convolution operations only on part of the channels to reduce the model’s parameters. Next, the Spatial Pyramid Pooling module is improved by using convolutions with different dilation rates to enhance the extraction of fine-grained features. Then, GLFPN is proposed by combining the Global-to-Local Spatial Aggregation (GLSA) module with the improved BiFPN structure. Its weighted fusion can more effectively retain small sample features, thereby improving the model’s accuracy. After that, a more lightweight detection head is designed to further reduce the model’s parameters and computational load. Finally, the WIoU loss function is used to replace the original CIoU. Experimental results show that the improved model achieves an mAP50 of 79.6% on the NEU-DET dataset, a 4.2 percent improvement over YOLOv8n, while the model parameters and computational load are only 1.43×10⁶ and 4.7×10⁹, respectively, representing a 53.3% and 41.9% reduction compared to YOLOv8n. This makes the model more accurate while being more suitable for deployment on edge devices.