During the detection of wire breakage defects in the constant-diameter growth process of monocrystalline silicon production,an improved YOLOv8-RCA network framework was proposed based on the YOLOv8 object detection model. First,the RFAConv convolution was introduced to enhance the model’s ability to capture features at different scales by incorporating receptive field attention. Second,a novel C2f_Atten module was designed with the addition of CI_Attention,which enabled the model to focus more effectively on key regions during detection and aimed to improve accuracy. Furthermore,the AFGCAttention mechanism was introduced to enhance the diversity of features input into the attention head and reduce the computational redundancy of multi-head attention. Experimental results show that YOLOv8-RCA achieved excellent performance in multiple aspects. The Precision,Recall,mAP50,and mAP50-95 increase by 12.6%,6.2%,8.1%,and 5.6%,respectively,compared with the baseline model. The model was tested and verified using a constant-diameter growth process wire breakage dataset and is applicable to actual monocrystalline silicon production.
ZHAO Shun-xiang, ZHOU An-liang, WANG Jia, WEN Meng, LIN Yu-long, CHANG Gao-yu, HOU Shu-yu, ZHANG Di.
Defect Detection in Constant-diameter Growth Process of Monocrystalline Silicon Based on an Improved YOLOv8n Model[J].
Science Technology and Engineering, 2026, 26(4): 1565-1574 DOI:10.12404/j.issn.1671-1815.2500354
现阶段,目标检测算法主要基于深度学习模型,主要形式可分为两种:单阶段的目标检测算法和双阶段的目标检测算法。单阶段算法无需候选区域(region proposal),能在单一阶段直接输出物体的类别概率和位置坐标,如SSD(single shot multiBox detector)[8]、YOLO(you only look once)[9-14]系列等。这类算法结构简单,相较于双阶段算法,更易于实现。双阶段算法优先生成候选框,之后对这些候选框进行目标的识别以及定位,因此检测精度相比较单阶段较高。然而,由于需要先筛选候选框,双阶段算法的检测速度较慢。因此,在对实时性要求较高的场景中,单阶段算法能够更好地满足需求。
YOLOv8(you only look once version 8)[15]是Ultralytics公司在2023年1月发布的一代实时目标检测模型。其结构由骨干网络、颈部网络、检测头组成,采用无锚点的头部设计,这样做的目的是在保证检测准确性的同时对检测速度进一步提升。YOLOv8沿用了YOLO系列算法的核心思想:将目标检测任务转化为回归问题,直接从图像中预测边界框和类别标签。其典型结构包括:特征提取网络(backbone)、特征融合网络(Neck)以及对应的检测头(Head),网络结构图如图1所示。其中特征提取网络负责从输入图像中提取有用的信息,并将提取的特征传入下一层网络。特征融合网络负责将不同尺度的输入特征进行细致融合,以此来应对不同大小的检测目标,其能够同时检测小(small)、中(medium)、大(large)3种尺度的目标。最后检测头是网络的输出部分,负责生成最终的检测结果,其采用多尺度预测的方式,每个尺度一个检测头,来应对检测的鲁棒性。
基于YOLOv8n,实现对模型进行网络优化,提出一种断线缺陷检测算法YOLOv8-RCA,其网络结构如图3所示。主要工作包含以下几个方面:①针对输入图片多、类型复杂等问题,创新改进C2f模块,加入CI_Attenton模块,使其能够区分不同背景的输入图片,对检测目标有着更好的识别效果,能够有效的提高准确率;②针对多类型检测精度低的问题,引入了RFAConv[16]模块,能够更好地调整感受野来提高检测精度,提高网络在处理复杂场景或多尺度特征时的表现;③采用AFGCAttention[17]注意力机制,实现图像去雾的特征,模块在小目标检测头之前,对小目标有着更好的检测能力;④使用CIOU(Complete Intersection over Union)[18]作为边界框损失函数,通过综合考虑重叠区域、中心点距离和宽高比,实现更快的收敛速度。创新性地融合了RFAConv卷积、C2f_Atten模块以及AFGCAttention注意力机制等多个改进模块于YOLOv8网络框架中。这种多模块协同的改进方式,全方位地提升了模型对于单晶硅等径工艺缺陷的检测能力,包括对不同尺度特征的捕捉、关键区域的关注以及特征多样性的增强等多方面,使得模型在复杂工业场景下的检测性能得到了显著优化,为工业缺陷检测领域提供了全新的技术思路。
在深度学习模型中,无论是backbone,还是Neck中都使用到卷积层模块(Conv),当然Conv是负责提取特征的关键模块。然而,传统卷积操作对卷积的位置和权重是固定的,缺乏网络对输入图片特征的动态自适应能力,尤其在处理多样分类问题或者是复杂的非均匀分布特征问题中,表现受到限制。具体表现为,固定的感受野限制了卷积层捕捉图像全局上下文信息的灵活性,与可变卷积核或扩张卷积相比,其在处理复杂背景或大尺寸目标图像时存在不足。较小的卷积核可能无法全面覆盖目标特征,而较大的卷积核则可能引入过多无关信息,干扰模型对目标的精准识别。为了增强网络对局部特征的捕捉能力和适配性,引入一种卷积操作RFAConv(receptive field adaptive convolution)模块,该模块通过感受野注意力RFA(receptive field attention)机制的加入,其主要目的是解决根据输入的特征动态调整感受野大小和权重分配,同时也考虑感受野中每个特征的重要性问题,增强模型对复杂图像的处理能力[19]。该方法能够解决了传统卷积操作中参数共享问题,卷积核为3×3的RFAConv结构如图7所示。
对YOLOv8-RCA算法从以下4个方面进行评估,分别是准确率(Precision,记为P)、召回率(Recall,记为R)、IoU阈值为0.5的平均精度均值(mean average precision,mAP50)、IoU为0.5~0.95(其中步长为 0.05)加权平均后的平均精度均值(mAP50-95)。其中平均精度均值定义为对所有类别的平均精度(average precision,AP)求平均值。
LiuLixin, LuoPing, LiChun, et al. Growth principle and technology of monocrystalline silicon[J]. Journal of Changchun University of Science and Technology, 2009, 32(4): 569-573.
WangZhengsheng, RenYongsheng, MaWenhui, et al. Growth principle, technology and prospect of Czochralcian silicon[J]. Materials Review, 2024, 38(9): 5-17.
[5]
CzochralskiJ. Ein neues verfahren zur messung der kristallisationsgeschwindigkeit der metalle[J]. Zeitschrift Für Physikalische Chemie, 1918, 92: 219-221.
[6]
李欣鸽. CZ硅单晶等径生长阶段关键参数优化研究[D]. 郑州: 郑州大学, 2020.
[7]
LiXinge. Study on optimization of key parameters of CZ silicon single crystal equidiameter growth stage[D]. Zhengzhou: Zhengzhou University, 2020.
WangLongye, HuangJun, ZengXiaoli. Bare board defect detection of printed circuit board based on YOLO-PCB[J]. Science Technology and Engineering, 2024, 24(15): 6338-6345.
YangZhou, ChengYing, ZhangShijing, et al. Research on Czochralski single crystal silicon dislocation detection model based on improved YOLOv5 algorithm[J]. Journal of Applied Optics, 2019, 44(5): 1022-1029.
ZhaiYongjie, XuWei, HanYuchen, et al. Digital instrument detection method based on ID-YOLO[J]. Science Technology and Engineering, 2019, 24(16): 6775-6782.
[14]
LiuW, AnguelovD, ErhanD, et al. Ssd: single shot multibox detector[C]// Computer Vision-ECCV 2016: 14th European Conference. Amsterdam: Springer International Publishing, 2016: 21-37.
[15]
LiC, LiL, JiangH, et al. YOLOv6: a single-stage object detection framework for industrial applications[J]. arXiv Preprint, arXiv: 2209.02976, 2022.
[16]
RedmonJ. Yolov3: an incremental improvement[J]. arXiv Preprint, arXiv: 1804.02767, 2018.
RedmonJ. You only look once: unified, real-time object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2016: 91.
[19]
SohanM, RamT S, ReddyC V R. A review on yolov8 and its advancements[C]// International Conference on Data Intelligence and Cognitive Informatics. Singapore: Springer, 2024: 239.
[20]
KhanamR, MuhammadH. Yolov11: an overview of the key architectural enhancements[J]. arXiv Preprint, arXiv: 2410.17725, 2024.
[21]
VargheseR, SambathM.YOLOv8: a novel object detection algorithm with enhanced performance and robustness[C]//2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS). New York: IEEE, 2024: 1-6.
[22]
ZhangX, LiuC, YangD, et al. RFAConv: innovating spatial attention and standard convolutional operation[J]. arXiv Preprint, arXiv: 2304.03198, 2023.
[23]
SunH, WenY, FengH, et al. Unsupervised Bidirectional contrastive reconstruction and adaptive fine-grained channel attention networks for image dehazing[J]. Neural Networks, 2024, 176: 106314.
[24]
ZhengZ, WangP, RenD, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Transactions on Cybernetics, 2021, 58: 8574-8586.