Efficient defects segmentation from photovoltaic (PV) electroluminescence (EL) images is a crucial process due to the random inhomogeneous background and unbalanced crack non-crack pixel distribution. The automatic defect inspection of solar cells greatly influences the quality of photovoltaic cells, so it is necessary to examine defects efficiently and accurately. In this paper we propose a novel end to end deep learning-based architecture for defects segmentation. In the proposed architecture we introduce a novel global attention to extract rich context information. Further, we modified the U-net by adding dilated convolution at both encoder and decoder side with skip connections from early layers to later layers at encoder side. Then the proposed global attention is incorporated into the modified U-net. The model is trained and tested on Photovoltaic electroluminescence 512x512 images dataset and the results are recorded using mean Intersection over union (IOU). In experiments, we reported the results and made comparison between the proposed model and other state of the art methods. The mean IOU of proposed method is 0.6477 with pixel accuracy 0.9738 which is better than the state-of-the-art methods. We demonstrate that the proposed method can give effective results with smaller dataset and is computationally efficient.
U-Net Based Defects Inspection in Photovoltaic Electroluminecscence Images
Rahman, Muhammad Rameez Ur;Xi, Wen
2019-01-01
Abstract
Efficient defects segmentation from photovoltaic (PV) electroluminescence (EL) images is a crucial process due to the random inhomogeneous background and unbalanced crack non-crack pixel distribution. The automatic defect inspection of solar cells greatly influences the quality of photovoltaic cells, so it is necessary to examine defects efficiently and accurately. In this paper we propose a novel end to end deep learning-based architecture for defects segmentation. In the proposed architecture we introduce a novel global attention to extract rich context information. Further, we modified the U-net by adding dilated convolution at both encoder and decoder side with skip connections from early layers to later layers at encoder side. Then the proposed global attention is incorporated into the modified U-net. The model is trained and tested on Photovoltaic electroluminescence 512x512 images dataset and the results are recorded using mean Intersection over union (IOU). In experiments, we reported the results and made comparison between the proposed model and other state of the art methods. The mean IOU of proposed method is 0.6477 with pixel accuracy 0.9738 which is better than the state-of-the-art methods. We demonstrate that the proposed method can give effective results with smaller dataset and is computationally efficient.I documenti in ARCA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.