| Title: | About the Application of Autoencoders For Visual Defect Detection |
| Authors: | Rádli, Richárd Czúni, László |
| Citation: | WSCG 2021: full papers proceedings: 29. International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision, p. 181-188. |
| Issue Date: | 2021 |
| Publisher: | Václav Skala - UNION Agency |
| Document type: | konferenční příspěvek conferenceObject |
| URI: | http://hdl.handle.net/11025/45023 |
| ISBN: | 978-80-86943-34-3 |
| ISSN: | 2464-4617 2464–4625(CD/DVD) |
| Keywords: | neuronová síť autoencoderu;konvoluční neuronová síť;detekce defektů;detekce anomálií bez dozoru |
| Keywords in different language: | autoencoder neural network;convolutional neural network;defect detection;unsupervised anomaly detection |
| Abstract in different language: | Visual defect detection is a key technology in modern industrial manufacturing systems. There are many possibleappearances of product defects, including distortions in color, shape, contamination, missing or superfluous parts.For the detection of those, besides traditional image processing techniques, convolutional neural networks basedmethods have also appeared to avoid the usage of hand-crafted features and to build more efficient detectionmechanisms. In our article we deal with autoencoder convolutional networks (AEs) which do not require examplesof defects for training. Unfortunately, the manual and/or trial-and-error design of AEs is still required to achievegood performance, since there are many unknown parameters of AEs which can greatly influence the detectionabilities. For our study we have chosen a well performing AE known as structural similarity AE (SSIM-AE),where the loss function and the comparison of the output with the input is implemented via the SSIM instead ofthe often used L1 or L2 norms. Investigating the performance of SSIM-AE on different data-sets, we found that itsperformance can be improved with modified convolutional structures without modifying the size of latent space.We also show that finding a model with low reconstruction error during training does not mean good detectionabilities and denoising AEs can increase efficiency. |
| Rights: | © Václav Skala - UNION Agency |
| Appears in Collections: | WSCG 2021: Full Papers Proceedings |
Please use this identifier to cite or link to this item:
http://hdl.handle.net/11025/45023Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.