C2DEM-YOLO: improved YOLOv8 for defect detection of photovoltaic cell modules in electroluminescence image

Photovoltaic (PV) cell modules are the core components of PV power generation systems, and defects in these modules can significantly affect photovoltaic conversion efficiency and lifespan. Electroluminescence (EL) testing is a method used to detect defects during the production process of these mod...

Full description

Saved in:
Bibliographic Details
Published in:Nondestructive testing and evaluation 2025-01, Vol.40 (1), p.309-331
Main Authors: Zhu, Jiahao, Zhou, Deqiang, Lu, Rongsheng, Liu, Xu, Wan, Dahang
Format: Article
Language:English
Subjects:
Accuracy
attention mechanism
auxiliary regression boxes
Datasets
defect detection
Defects
Electroluminescence
Energy conversion efficiency
Feature extraction
Modules
Object recognition
Photovoltaic cell modules
Photovoltaic cells
Photovoltaic conversion
Regression models
Solar power generation
YOLOv8
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Photovoltaic (PV) cell modules are the core components of PV power generation systems, and defects in these modules can significantly affect photovoltaic conversion efficiency and lifespan. Electroluminescence (EL) testing is a method used to detect defects during the production process of these modules. To address the issue of low defect detection accuracy caused by the complex background and large-scale variations of EL images, we propose an object detection network named C2DEM-YOLO to improve the accuracy of defect detection. Firstly, a deep-shallow feature extraction module called C2Dense is designed to replace the C2f module in the YOLOv8's backbone. Secondly, a cross-space multi-scale attention(EMA) is introduced after C2Dense to apply pixel-level attention to the extracted features, which suppresses background information while enhancing useful features for defect detection. Finally, by replacing CIoU with Inner-CIoU, we introduce auxiliary regression boxes to improve the accuracy of detection and the generalisation ability of the model. Experimental results show that C2DEM-YOLO achieves an average precision of 92.31% on the PVEL-AD dataset, which has 2.41%, 1.93%, and 1.56% improvement compared to YOLOv5s, YOLOv8n, YOLOv8s, respectively. Moreover, on our self-built dataset, the mAP@0.5 and mAP@0.5:0.95 of C2DEM-YOLO are improved by 1.42% and 1.46% compared to YOLOv8n, reaching 84.07%.
ISSN:1058-9759
1477-2671
DOI:10.1080/10589759.2024.2319263