Anomaly detection in industrial vision with convolutional autoencoders
(2024)
Files
ThesisASC.pdf
Open Access - Adobe PDF
- 31.66 MB
Details
- Authors
- Supervisors
- De Vleschouwer, Christophe
- Abstract
- The identification of defective samples in production lines is a critical challenge in the field of industrial vision. Due to the scarcity and variability of anomalies, it is assumed that only normal images are available during training. This limitation renders supervised methods unsuitable and necessitates the design of an unsupervised approach that addresses the problem in a less direct formulation. In this work, we utilize deep learning models to localize potential defects in objects based on a single image. The fundamental principle underlying our methodology is the capacity of a convolutional autoencoder to reconstruct a defect-free version of an input image. Defective structures are then associated with regions exhibiting high reconstruction residuals. However, a significant limitation of this approach is its reduced effectiveness in detecting defects that exhibit low contrast with their surroundings. To address this, two extensions have been introduced with the objective of enhancing detection performance, particularly for low-contrast defects. These extensions are driven by in-depth visualizations of the autoencoder's internal representations. This thesis presents a comprehensive analysis of the newly introduced approaches, both qualitatively and quantitatively. The study is conducted over the MVTec AD dataset, which reflects a practical industrial use case. However, the objective is to develop a comprehensive understanding of the properties of these methods, ensuring their applicability to a wide range of industrial use cases.
- Affiliations
Citations
Collin, A.-S. (2024). Anomaly detection in industrial vision with convolutional autoencoders. https://hdl.handle.net/2078.5/247430