Attention M-net for Automatic Pixel-Level Micro-crack Detection of Photovoltaic Module Cells in Electroluminescence Images

In the solar power industry, quality inspection of solar cells is a very important part of the production and application process. Micro-crack is a type of common defect that may be present in photovoltaic (PV) module cells which can reduce the power generation efficiency and service lifetime. However, it is difficult to identify micro-crack by the naked eye because it is readily confused with non-uniform backgrounds and complex textures in Electroluminescence (EL) images. In this paper, we propose Attention M-net which combines efficient segmentation model structure and attention mechanism. It is a novel micro-crack detection model for automated pixel-level micro-crack detection of PV module cells. The M-shaped structure solves "All Black" issue that is easy to occur due to the severe imbalance of the micro-crack segmentation dataset. And integration of attention mechanism into the network significantly improves the accuracy of segmentation. Because of the above two advantages, the proposed model can be accurately learned from a small annotated dataset, thereby saving the time of micro-crack EL images collection and pixel-level annotations. We extensively trained the proposed model on a small dataset of 20 annotated EL images which include 10 monocrystalline and 10 polycrystalline. The evaluated results on the test dataset demonstrate the high efficiency and accuracy of Attention M-net for pixel-level micro-crack detection of PV module cells in EL images.