摘要
针对光伏组件表面缺陷检测任务中,现有深度学习模型复杂度高、实时性不足的问题,提出一种融合轻量化结构与分级知识蒸馏的光伏电池表面缺陷检测算法。该算法在CenterNet网络的基础上,设计双路径特征金字塔网络,通过双向交互式特征传递机制有效提高特征融合效率。然后提出改进的高斯核生成策略,通过以中心点为基准的放射性高斯核覆盖方式,优化了网络的损失函数,有效提升网络对缺陷区域的定位敏感度。在此基础上,本文进一步提出分级知识蒸馏方案,通过两阶段迁移体系,从复杂到简化逐步传递知识,有效解决了教师-学生网络架构差异带来的知识迁移低效问题。实验结果显示,所提出算法的平均精度达到了83.7%,同时将推理模型的参数量减少至5.7 M,实时检测速率可达143.4帧每秒,兼顾检测精度与实时性需求,满足光伏产线在线检测要求。
To address the issues of high complexity and insufficient real-time performance in existing deep learning models for photovoltaic(PV)module surface defect detection,this paper proposes a surface defect detection algorithm integrating lightweight architecture and hierarchical knowledge distillation.Based on the CenterNet framework,the algorithm introduces a dual-path feature pyramid network with a bidirectional interactive feature which could transfer mechanism to enhance feature fusion efficiency.An improved Gaussian kernel generation strategy is further developed,which employs a radiation-based Gaussian kernel coverage centered on target points to optimize the loss function,thereby increasing the network's sensitivity to defect localization.Additionally,a hierarchical knowledge distillation scheme is proposed,enabling a two-stage knowledge transfer from complex to simplified models to resolve inefficiencies caused by architectural discrepancies between teacher and student networks.Experimental results demonstrate that the proposed algorithm achieves an average precision of 83.7%,and model parameters are reduced to 5.7M,and a real-time detection speed of 143.4 frames per second(FPS)is attained.This approach effectively balances detection accuracy and computational efficiency,meeting the requirements for online inspection in PV production lines.
作者
梁龙学
袁桂花
吴小所
闫浩文
LIANG Longxue;YUAN Guihua;WU Xiaosuo;YAN Haowen(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《兰州交通大学学报》
2025年第4期105-116,共12页
Journal of Lanzhou Jiaotong University
基金
国家重点研发计划(2022YFB3903604)
甘肃省科技计划项目(24JRZA104)。
关键词
光伏电池
缺陷检测
深度学习
特征金字塔
知识蒸馏
photovoltaic cells
defect detection
deep learning
feature pyramid network
knowledge distillation
