Photovoltaic panel defect detection algorithm based on infrared imaging and improved YOLOv8

Abstract

To address the challenges of high missed detection rates, complex backgrounds, unclear defect features, and uneven difficulty levels in target detection during the industrial process of photovoltaic panel defect detection, this article proposes an infrared detection method based on computer vision, with enhancements built upon the YOLOv8 model. First, a multi-channel squeeze-and-excitation network is introduced to improve feature extraction capabilities and is integrated into the neck network. Second, GhostConv and BoTNet are incorporated into the backbone network to reduce model parameters while enhancing defect detection performance. Finally, the Focaler-Complete Intersection over Union (Focaler-CIoU) loss function is employed to tackle the issue of imbalanced difficulty in target detection tasks. The method is evaluated on the PV-Multi-Defect-main dataset and further validated through a generalization test on the PVEL-AD dataset. Results demonstrate that, compared with the baseline YOLOv8 model, the proposed approach achieves significant improvements in precision (3.6%), recall (10.4%), mAP50 (4.8%), and mAP50–95 (4.5%) while maintaining nearly the same parameter count. On the PVEL-AD dataset, the method effectively addresses the challenge of feature extraction failure for dislocation-type defects, achieving substantial gains in precision (7.8%), recall (17.1%), mAP50 (19.5%), and mAP50–95 (13.2%). Furthermore, comparisons with several state-of-the-art detection algorithms reveal that the proposed method consistently delivers improved detection performance, validating its effectiveness as a robust solution for photovoltaic panel defect detection.