摘要
针对现有目标检测方法在光伏电池电致发光图像中存在的小目标漏检率高、复杂背景干扰鲁棒性不足及跨尺度缺陷检测能力有限等问题,提出基于MFES-YOLOV8n的缺陷检测模型,旨在提升工业场景下的检测精度与效率。首先,在主干网络中嵌入C2f-ST特征提取模块,通过Swin Transformer的窗口自注意力机制,增强微小缺陷的局部-全局特征关联,结合残差连接保留浅层细节特征,提升细粒度特征提取能力;其次,设计ES-SPPCSPC特征表达模块,融合群卷积与增强型SimAM注意力机制,通过能量基、通道和空间三重注意力协同优化,动态抑制背景噪声,增强缺陷特征特异性;最后,构建MSFF-Neck多尺度特征融合模块,采用尺度序列特征融合和三重特征编码策略,实现深层语义与浅层细节的互补交互,缓解多尺度特征衰减问题。实验在PVEL-AD数据集上验证了模型的有效性,结果表明,该模型以6.1 M参数量达到0.897的mAP@0.5,较基准模型YOLOv8n提升3.0%。本研究通过“细粒度特征提取—跨尺度语义增强—多层级特征融合”的递进式优化,突破了传统模型在多类别跨尺度缺陷检测中的性能瓶颈,为工业场景提供了高精度、轻量化且适配边缘计算的缺陷检测方案,在维持低计算复杂度的同时,满足工业场景对实时性与可靠性的要求,为推动光伏产业质量控制与智能化运维提供了技术支持。
To address the issues of high missed detection rates for small targets,insufficient robustness against complex background interference,and limited cross-scale defect detection capabilities in existing target detection methods for photovoltaic cell electroluminescence images,a defect detection model based on MFES-YOLOV8n is proposed to enhance detection accuracy and efficiency in industrial scenarios.First,a C2f-ST feature extraction module is embedded into the backbone network,utilizing the window-based self-attention mechanism of Swin Transformer to strengthen local-global feature associations for micro-defects,combined with residual connections to preserve shallow-layer detail features,thereby improving fine-grained feature extraction capabilities.Second,an ES-SPPCSPC feature representation module is designed,integrating group convolution with an enhanced SimAM attention mechanism,achieving dynamic suppression of background noise and enhancement of defect-specific features through synergistic optimization of energy-based,channel,and spatial attention.Finally,an MSFF-Neck multi-scale feature fusion module is constructed,employing scale-sequential feature fusion and triple feature encoding strategies to enable complementary interactions between deep semantic and shallow detail features,mitigating multi-scale feature degradation.Experiments on the PVEL-AD dataset validate the model′s effectiveness,demonstrating that it achieves an mAP@0.5 of 0.897 with 6.1 M parameters,improving by 3.0%over the baseline YOLOv8n.Through a progressive optimization strategy of“fine-grained feature extraction,cross-scale semantic enhancement,and multi-level feature fusion,”this study overcomes performance bottlenecks in multi-category and cross-scale defect detection of traditional models,providing a high-precision,lightweight,and edge-computing-compatible defect detection solution for industrial scenarios.While maintaining low computational complexity,it meets the demands for real-time performance and reliability in industrial applications,offering technical support for advancing quality control and intelligent maintenance in the photovoltaic industry.To address the issues of high missed detection rates for small targets,insufficient robustness against complex background interference,and limited cross-scale defect detection capabilities in existing target detection methods for photovoltaic cell electroluminescence images,a defect detection model based on MFES-YOLOV8n is proposed to enhance detection accuracy and efficiency in industrial scenarios.First,a C2f-ST feature extraction module is embedded into the backbone network,utilizing the window-based self-attention mechanism of Swin Transformer to enhance local-global feature associations for micro-defects,combined with residual connections to preserve shallow-layer detail features.Therefore,the fine-grained feature extraction capabilities are improved.Secondly,an ES-SPPCSPC feature representation module is designed,integrating group convolution with an enhanced SimAM attention mechanism,achieving dynamic suppression of background noise and enhancement of defect-specific features through synergistic optimization of energy-based,channel,and spatial attention.Finally,an MSFF-Neck multi-scale feature fusion module is established,employing scale-sequential feature fusion and triple feature encoding strategies to enable complementary interactions between deep semantic and shallow detail features,mitigating multi-scale feature degradation.Experiments on the PVEL-AD dataset validate the model’s effectiveness.Results show that it achieves an mAP@0.5 of 0.897 with 6.1 M parameters,improving by 3.0%over the baseline YOLOv8n.Through a progressive optimization strategy of“fine-grained feature extraction,cross-scale semantic enhancement,and multi-level feature fusion”,this study overcomes performance bottlenecks in multi-category and cross-scale defect detection of traditional models,providing a high-precision,lightweight,and edge-computing-compatible defect detection solution for industrial scenarios.While maintaining low computational complexity,it meets the demands for real-time performance and reliability in industrial applications,offering technical support for advancing quality control and intelligent maintenance in the photovoltaic industry.
作者
陈俊生
陈沂蒙
刘明杰
朴昌浩
Chen Junsheng;Chen Yimeng;Liu Mingjie;Piao Changhao(College of Automation,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《仪器仪表学报》
北大核心
2025年第6期251-262,共12页
Chinese Journal of Scientific Instrument
基金
国家重点研发计划(2022YFE0101000)项目资助。
