To address the issue of missed and false detections of steel surface defects under conditions of complex textures, weak contrast, and edge deployment, a detection algorithm oriented to edge deployment named QHN-YOLO was proposed. Firstly, a quantization-friendly convolution (QFConv) was constructed to replace the original convolution module, so as to stabilize activation distribution and improve edge deployment efficiency. Secondly, a hybrid dilated residual attention block (HDRAB) was embedded into the C3k2 module to form C3k2_HDRAB, which was used to enhance multi-scale feature extraction and small defect representation ability. Finally, a noise-variance prior detection head (NVPD) was designed to suppress background interference and improve defect response discriminability. Experimental results indicate that the mAP of the improved model reaches 80.2%, with 2.8×106 parameters and 7.1 GFLOPs of floating-point computations. Under edge deployment conditions, the model inference latency is 28.9 ms, and the detection speed reaches 34.6 fps, indicating real-time detection capability in industrial scenarios. The improved model balances detection accuracy and lightweighting and improves the edge application performance of steel surface defect detection.
式中: X 表示输入特征图; Y 表示卷积输出;为卷积核权重; Z 为归一化后输出的特征图;为偏置;为输入通道数;h,w为特征图空间尺寸;为卷积核大小;与分别为输入与输出通道索引;与分别表示卷积核在高度方向和宽度方向上的空间索引;与分别为均值与方差;与为可学习缩放与平移参数;为数值稳定项;表示逐通道乘法;为最终激活输出的特征图.
为了验证本文方法的有效性,本次实验采用的评估指标包括精确率(P)、召回率(R)、平均精确率(AP)、平均精确率均值(mAP)、参数量、计算量和检测速度(v).其中计算量用浮点运算次数(floating point operations,FLOPs)表示,GFLOPs指模型需要进行109次浮点运算;检测速度用每秒帧率(fps)表示,即模型每秒钟处理的图像帧数.各指标计算公式为
SchlegelJ. The world of steel [M]. Wiesbaden: Springer, 2023: 45-49.
[2]
PalitP, SriK P, KayalN, et al. Failure investigation and crack characterization of spalled work roll in hot strip mill [J]. Journal of Failure Analysis and Prevention, 2024, 24(4): 1522-1532.
[3]
DwivediS K, VishwakarmaM, SoniA. Advances and researches on non destructive testing: a review [J]. Materials Today: Proceedings, 2018, 5(2): 3690-3698.
[4]
SaberironaghiA, RenJ, El-GindyM. Defect detection methods for industrial products using deep learning techniques: a review [J]. Algorithms, 2023, 16(2): 95.
[5]
BaiJ, WuD, ShelleyT, et al. A comprehensive survey on machine learning driven material defect detection [J]. ACM Computing Surveys, 2025, 57(11): 1-36.
[6]
AlhashmiS A, Al-azawiA. A review of the single-stage vs. two-stage detectors algorithm: comprehensive insights into object detection [J]. International Journal of Environmental Sciences, 2025, 11(3): 775-787.
[7]
BharatiP, PramanikA. Deep learning techniques—R-CNN to mask R-CNN: a survey [C]// Computational Intelligence in Pattern Recognition (CIPR). Singapore: Springer, 2020: 657-668.
[8]
RenS Q, HeK M, GirshickR, et al. Faster R-CNN: towards real-time object detection with region proposal networks [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[9]
HeK M, GkioxariG, DollárP, et al. Mask R-CNN [C]// 2017 IEEE International Conference on Computer Vision (ICCV). Venice, 2017: 2980-2988.
[10]
KangS Z, HuZ Y, LiuL J, et al. Object detection YOLO algorithms and their industrial applications: overview and comparative analysis [J]. Electronics, 2025, 14(6): 1104.
[11]
AnY S, ChenX, HeX, et al. FSU-YOLO: aero-engine blade defect detection based on improved YOLOv8 [J]. Measurement Science and Technology, 2025, 36(10): 105417.
[12]
LiJ, HeX, ChenX, et al. HDFA-YOLO: a real-time steel surface defect detection model based on backbone lightweight design and multi-scale feature fusion [J]. Measurement, 2026, 258: 119390.
HaoBo, XuXin-yan, ZhaoYu-xin, et al. Surface defect detection of riveting holes based on improved YOLOv8 [J]. Journal of Northeastern University (Natural Science), 2024, 45(11): 1595-1603.
[23]
KhanamR, HussainM. YOLOv11: an overview of the key architectural enhancements [EB/OL]. (2024-10-23)[2025-12-25].
[24]
YuJ Y, ShuC Y, ZhouS F, et al. FQ-PETR: fully quantized position embedding transformation for multi-view 3D object detection [EB/OL]. (2025-11-12)[2025-12-09].
HeY, SongK C, MengQ G, et al. An end-to-end steel surface defect detection approach via fusing multiple hierarchical features [J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(4): 1493-1504.
[27]
RedmonJ, FarhadiA. YOLOv3: an incremental improvement [EB/OL]. (2018-04-08)[2025-12-25].
[28]
AdarshP, RathiP, KumarM. YOLOv3-tiny: object detection and recognition using one stage improved model [C]// 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS). Coimbatore, 2020: 687-694.
WangC Y, YehI H, MarkL H. YOLOv9: learning what you want to learn using programmable gradient information [C]// European Conference on Computer Vision (ECCV). Cham: Springer, 2025: 1-21.
[33]
ChenH, ChenK, DingG G, et al. YOLOv10: real-time end-to-end object detection [C]// Advances in Neural Information Processing Systems. Vancouver, 2024: 107984-108011.
SelvarajuR R, CogswellM, DasA, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization [J]. International Journal of Computer Vision, 2020, 128(2): 336-359.