Abstract
ABSTRACT One of the challenges in the field of photovoltaics (PV) is the automation of defect detection in electroluminescent (EL) images of PV cells. This is due to the similarities between defects and the intricate nature of the background, which can make it difficult to accurately identify and distinguish defects. In response to this problem, we introduce the Efficient Long-Range Convolutional Network (ELCN) module, designed to enhance defect detection capabilities in EL images of PV cells. The ELCN module is based on the ConvNeXt block, renowned for its efficiency and scalability, and integrates the design principles of the Cross-Stage Partial Network (CSPNet). This unique design facilitates a higher level of gradient combination while simultaneously reducing computational overhead. By incorporating the ELCN into the YOLOv7 object detector, we create a novel end-to-end ELCN-YOLOv7 framework, improving accuracy and reducing model parameters for detecting defects in raw EL images. Furthermore, to boost the accuracy of ELCN-YOLOv7 even further, we propose a two-stage fine-tuning method. This approach leverages similar small datasets to assist in the fine-tuning process. On the PVEL-AD dataset, we validated the effectiveness of our proposed ELCN-YOLOv7 method. It achieved a 91.93% mAP and 94.34 FPS, representing improvements of 3.19% points in mAP and 16.82 in FPS over the baseline YOLOv7 model. Additionally, our method outperforms previous approaches in both speed and accuracy, thereby establishing a new state-of-the-art performance.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have