Crack Inspection Using Convolutional Neural Networks and Information Entropy Data Fusion

The crack defect inspection of Electroluminescence Images (EL) is a vital technique to avoid defective products entering the next photovoltaic module production process. However, the inhomogeneous texture background, disturbance of crystal grains pseudo defects and low contrast between crack and complex background are huge challenges for inspecting cracks. Thus, this paper proposes a novel deep learning architecture based on a convolutional neural network (CNN) and an Information entropy data fusion, called Entropy-CNN, to analyze EL images for crack inspection. In order to make full use of the spatial structure information of the crack to guide crack defect inspection, a novel data fusion strategy is proposed to aggregate information extracted from each surrounding images of the target crack image. The data fusion strategy discards false positives effectively caused by crystal grains pseudo defects from the high-resolution images with complex random backgrounds in actual production line, which significantly enhanced the overall performance of the system. The Entropy-CNN achieves impressively high performance on inspecting crack defects in EL images with a precision of 95.7%. The robustness and adaptability of the proposed Entropy-CNN are further tested on 65 images of 960 ×960 pixel resolutions under various conditions (e.g., different intensity lighting and diverse cracks).