Abstract:
Traditional methods for defect detection applied in industry are complex, time-consuming, not robust and demanding for professional experience due to hand-crafted feature...Show MoreMetadata
Abstract:
Traditional methods for defect detection applied in industry are complex, time-consuming, not robust and demanding for professional experience due to hand-crafted features extraction and pipeline design. Besides, current deep learning based methods for general object segmentation demand for a large number of region-level human annotations.Instead, we present DefectGAN for defect detection in a weakly-supervised learning, which requires very a few human annotations. In practical application, images in training dataset are merely labeled with two categories: negative and positive. Despite being trained on image-level rather than region-level labels, DefectGAN has remarkable ability of localizing defect regions.DefectGAN can have comparable and visually even better performance than SegNet, a supervised learning method on dataset CCSD-NL and DAGM 2007. The detected regions are more similar to the original defect regions visually and it has the potential of detecting unseen defects.
Date of Conference: 22-26 August 2019
Date Added to IEEE Xplore: 19 September 2019
ISBN Information: