Cracks on sheet metals can significantly affect the overall strength. Crack detection during manufacturing is, thus, an important process for the quality assessment on a press line. Deep learning, a data-driven structure, has been extensively used to detect cracks on various surfaces. In this study, a crack detection technique for a press line using Retina Net and a novel data augmentation method is proposed, which mainly focuses on three steps, shape acquisition, style transfer, and edge fusion. First, the shapes of crack on different materials are extracted. Then, images are created by providing metal crack textures to those shapes using a fusion network with a relatively small number of real crack images. Real crack images are captured from a sheet metal forming line. Training data can be enriched using the proposed data augmentation method. Validation experiments are conducted to demonstrate the effectiveness of the proposed crack detection and data augmentation techniques.
|