Datenbestand vom 13. Juni 2019
Tel: 089 / 66060798
Mo - Fr, 9 - 12 Uhr
Fax: 089 / 66060799
DER VERLAG IST IN DER ZEIT VOM 12.06.2019 BIS 23.06.2019 AUSCHLIESSLICH PER EMAIL ERREICHBAR.
aktualisiert am 13. Juni 2019
978-3-8439-1387-4, Reihe Informatik
Automatic Visual Rope Inspection
193 Seiten, Dissertation Friedrich-Schiller-Universität Jena (2013), Softcover, A5
The automatic visual inspection of wire ropes and the associated problem of surface defect recognition is a challenging problem - even for the human. Anyhow, it is of great relevance for an enhancement of the existing inspection techniques. Recently developed inspection systems allow for a computer-assisted, automatic analysis. The main challenges with respect to this task arise from the diversity of the natural rope appearance, the unpredictable defect characteristics and the lack of defective training examples. Thus, supervised classification methods are not suitable. This work explores possible solutions to these problems by applying the concept of anomaly detection and closely related methods for one-class classification to the task of visual rope inspection.
The first part of the work deals with purely appearance-based approaches for anomaly detection. In the first instance, well-established features from the field of texture analysis are analyzed with respect to their suitability. We use them in combination with standard methods for one-class classification, such as clustering techniques, support vector methods or Gaussian processes to detect rope regions with an abnormal, suspicious appearance. In addition, we clarify the importance of context knowledge for the characterization of an anomaly. Besides this, sequential approaches are developed, which allow for a context-based analysis of the deviations with respect to their defect potential.
In the second part of the work, a model-based approach for anomaly detection in wire ropes is presented. A parametric 3d rope model and analysis-by-synthesis are used to compute artificial projections of the rope, which can be compared to the original rope images. On the one hand, this allows for an automatic estimation and monitoring of important rope parameters such as the lay lengths of strands and wires. Additionally, the alignment of rope model and real rope images facilitates the learning of the unknown relation between rope geometry and observed surface appearance. Given enough correspondences between geometry and appearance, this relation can be trained with a statistical model. In the following, we are able to determine the conditional likelihood of each gray value in the real rope image given the associated rope geometry.
All methods are evaluated on real-world rope data. The high defect detection accuracy can be considered as proof-of-concept and highlights the practical relevance of the work.