Detection of Cracks in Electroluminescence Images by Fusing Deep Learning and Structural Decoupling

Due to the heterogeneous texture background of solar cells, crack inspection becomes a challenging task. What's more, deep learning model is easily deceived by background interference, causing false detection. Thus a novel model called SEF-CNN is proposed to tackle the problem by exploring the function of traditional filter in deep learning approaches for crack defect detection, wherein the input images are filtered by steerable evidence filter (SEF), making more discriminative and robust features are obtained by convolutional neural networks (CNN).