Semantic segmentation method of photovoltaic cell microcracks based on EL polarization imaging

Abstract

Common defects arising from the production and use of cracked photovoltaic (PV) cells significantly impact both the power generation efficiency and service life of these cells. However, due to the complex environmental conditions at engineering sites and the small scale of microcracks, these defects are often confused with scratches or impurities on the surface of PV cells, leading to low detection efficiency and insufficient accuracy. This makes timely maintenance difficult. To address this issue, this thesis proposes a semantic segmentation method for PV cells based on polarization imaging, derived from the underlying mechanism of PV cell crack formation. In this approach, polarization imaging technology is utilized to capture three single-channel images of PV cells: polarization intensity (I), degree of polarization (DOP), and polarization quadrature (Q). These images are then stacked into different channels to create a comprehensive dataset. The dataset is annotated for training purposes, followed by an analysis of the texture features associated with microcrack defects. Additionally, the traditional U-Net architecture is enhanced to better exploit the unique features of the polarized three-channel images. Experimental results demonstrate that the proposed method outperforms existing semantic segmentation networks, showing excellent performance in microcrack defect segmentation in PV cells.