Pixel-Wise Defect Detection by CNNs without Manually Labeled Training Data. Issue 6 (12th May 2019)

• Record Type: Journal Article
• Authors: Haselmann, M.; Gruber, D. P.
• Is Part Of: Applied artificial intelligence. Volume 33:Issue 6(2019)
• Page Start: 548
• Page End: 566
• Publication Date: 2019-05-12
• Subjects: Artificial intelligence -- Periodicals; 006.3
• DOI: 10.1080/08839514.2019.1583862
• Languages: English
• ISSNs: 0883-9514

Abstract

ABSTRACT: In machine learning driven surface inspection one often faces the issue that defects to be detected are difficult to make available for training, especially when pixel-wise labeling is required. Therefore, supervised approaches are not feasible in many cases. In this paper, this issue is circumvented by injecting synthetized defects into fault-free surface images. In this way, a fully convolutional neural network was trained for pixel-accurate defect detection on decorated plastic parts, reaching a pixel-wise PRC score of 78% compared to 8% that was reached by a state-of-the-art unsupervised anomaly detection method. In addition, it is demonstrated that a similarly good performance can be reached even when the network is trained on only five fault-free parts.