School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi'an 710021, China
Show less
文章历史+
Received
Published
Revised
2025-04-07
2026-03-08
2025-12-10
Online Date
2026-07-09
PDF (15610K)
摘要
针对目前金属表面缺陷检测面临模型参数量大、精确度不足、工业现场部署困难等问题,提出了一种基于YOLOv8(you only look once version 8)的轻量化金属表面缺陷检测改进模型。该模型通过引入SCconv(spatial and channel reconst-ruction convolution)模块、C2f_Dual模块和倒残差多尺度注意力机制iEMA(inverted residual and efficient multi-scale attention),有效降低了模型参数量和计算量,同时提高了检测精度。实验使用SDI-YOLO在NEU-DET、Magnetic-Tile-Defect数据集上验证,mAP50分别比YOLOv8提高了2.5%、2.6%,参数量降低了15.6%,运算量减少了10.9%。与领域主流算法相比,SDI-YOLO在检测精度和参数量方面都有一定的优势,满足边缘设备部署中参数量较小、识别率高等需求。SDI-YOLO模型为金属表面缺陷检测提供了一种高效、轻量化的解决方案,具有重要的实际应用价值。
Abstract
To address the current challenges in metallic surface defect detection, such as large model parameters, insufficient precision, and deployment difficulties in industrial environments, an improved lightweight detection model based on YOLOv8 (you only look once version 8), named SDI-YOLO was proposed. The model integrated the spatial and channel reconst-ruction convolution module (SCconv), the C2f_Dual module, and the inverted residual and efficient multi-scale attention (iEMA) mechanism. These improvements effectively reduced the number of parameters and computational cost while enhanced detection accuracy. Experiments were conducted using SDI-YOLO on the NEU-DET and Magnetic-Tile-Defect datasets. The mAP50 values increased by 2.5% and 2.6%, respectively, compared with the YOLOv8. The number of parameters was reduced by 15.6%, and giga floating point operations per second was decreased by 10.9%.Compared with mainstream algorithms in the field, SDI-YOLO shows advantages in both detection accuracy and model size. It meets the requirements of edge device deployment, such as small parameter size and high recognition rate. SDI-YOLO provides an efficient and lightweight solution for metallic surface defect detection and has significant practical application value.
SHENG Liang-hao, JIA Xiao-yun, BAI Ying-jie, YANG Zhen-ying.
Lightweight Metal Surface Defect Detection Model Based on YOLOv8[J].
Science Technology and Engineering, 2026, 26(7): 2990-2999 DOI:10.12404/j.issn.1671-1815.2502382
Liu等[6]对YOLOv3(you only look once version 3)网络进行了重新设计,加入SARB(self-attention residual block)结构,并充分利用特征通道信息,但仍存在有效特征在深层网络中丢失问题突出,边缘提取能力较弱的问题。Jiang等[7]使用MobileNetv3和双向金字塔结构对深层次特征进行提取,引入K均值聚类算法 算法降低了网络损失,虽然提高了检测速度,但对于空间与通道的特征提取仍存在欠缺。张瑞芳等[8]引入注意力机制与跨尺度特征金字塔结构,以增强模型的特征提取能力,但模型的检测精度仍存在一定不足。周孟然等[9]通过引入FReLU(filter response-based rectified linear unit)激活函数和SIoU(shape-aware intersection over unit)损失函数,在空间解析方面取得了一定的进展,但模型存在参数量较大,帧率低等问题。
为客观评价模型性能并与基线模型进行全面对比,从模型体积、检测精度与运行效率等多个维度对比分析。使用:①参数量Para (parameters)衡量模型的整体复杂度和占用存储资源的程度;②精确率P (precision)表示预测为正样本的实例中实际为正样本比例,反映模型输出结果的可靠性;③召回率R (recall)表示正样本中检测到正样本的概率,用于衡量模型的漏检情况;④平均精度均值mAP50 (mean average precision at IoU=0.5) 用于衡量模型在所有类别上的平均识别精度,反映模型识别正样本的整体能力;⑤帧率 (frames per second,FPS)代表单位时间内可处理图像的帧数,衡量模型推理速度。具体公式为
YaoXirong. The impact of intelligent manufacturing on economic growth under the technological innovation-driven policy of Made in China 2025[J]. China Market, 2022(8): 1-4, 20.
[3]
LuoQ, FangX, LiuL, et al. Automated visual defect detection for flat steel surface: a survey[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(3): 626-644.
[4]
WangR, YuH, TangJ, et al. Optimal design of iron-cored coil sensor in magnetic flux leakage detection of thick-walled steel pipe[J]. Measurement Science and Technology, 2023, 34(8): 085123.
ZhouMengran, WangHaonan, GaoLipeng, et al. YOLOv5s-FCS based steel surface defect detection study[J]. Science Technology and Engineering, 2024, 24(14): 5901-5910.
[14]
LiJ, WenY, HeL. SCConv: spatial and channel reconstruction convolution for feature redundancy[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2023: 6153-6162.
[15]
ZhongJ, ChenJ, MianA. DualConv: dual convolutional kernels for lightweight deep neural networks[J]. IEEE Transactions onNeural Networks and Learning Systems, 2022, 34(11): 9528-9535.
[16]
OuyangD, HeS, ZhangG, et al. Efficient multi-scale attention module with cross-spatial learning[C]// Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2023). Piscataway: IEEE, 2023: 1-5.
[17]
ZhangJ, LiX, LiJ, et al. Rethinking mobile block for efficient attention-based models[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2023: 1389-1400.
[18]
WuY, HeK. Group normalization[C]// Proceedings of the European Conference on Computer Vision (ECCV). New York: IEEE, 2018: 3-19.
[19]
LiX, WangW, HuX, et al. Selective kernel networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2019: 510-519.
[20]
BaoY, SongK, LiuJ, et al. Triplet-graph reasoning network for few-shot metal generic surface defect segmentation[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-11.
[21]
HuangY, QiuC, YuanK. Surface defect saliency of magnetic tile[J]. The Visual Computer, 2020, 36(1): 85-96.
[22]
ShiX, ZhouS, TaiY, et al. An improved faster R-CNN for steel surface defect detection[C]// 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP). New York: IEEE, 2022: 1-5.
[23]
WangY, WangX, HaoR, et al. Metal surface defect detection method based on improved cascade R-CNN[J]. Journal of Computing and Information Science in Engineering, 2024, 24(4): 041002.
[24]
LiangC, WangZ Z, LiuX L, et al. SDD-Net: a steel surface defect detection method based on contextual enhancement and multiscale feature fusion[J]. IEEE Access, 2024, 11: 27-38.
[25]
GeQ, YuanH, ZhangQ, et al. Hyper-progressive real-time detection transformer (HPRT-DETR) algorithm for defect detection on metal bipolar plates[J]. International Journal of Hydrogen Energy, 2024, 74: 49-55.
[26]
LiJ, ChenM. DEW-YOLO: an efficient algorithm for steel surface defect detection[J]. Applied Sciences, 2024, 14(12): 5171.
[27]
ZhangH, FuW, WangX, et al. An efficient model for metal surface defect detection based on attention mechanism and multi-scale feature[J]. The Journal of Supercomputing, 2025, 81(1): 40.