The goal of this assignment is to investigate and implement new methods for visual quality control, based on images of products after they have been manufactured in 3D printers. Our previous approaches used flatbed scanners to obtain images of the surface of one particular product. These images were then manually checked for possible defects, thus obtaining an image data set suitable for machine learning. Several deep learning methods were evaluated for the fault detection.
The previous approach involved challenges, such as:
- Manual labelling is costly, so the dataset is relatively small
- Defects are relatively rare, yielding an imbalanced dataset
- The data is high-dimensional (images), but defects are typically only noticeable in a small area of the image
With a new data collection method, using cameras instead of a flatbed scanner, these challenges remain, and new ones arise:
- Images of the same product show more variety; factors such as location, orientation, and lighting conditions may vary
- 3D printing allows for customizability in products, previously unseen product shapes still need to be checked for defects in a reliable way
The suggested approach is to extend a previously tested approach that poses the problem as anomaly detection, using Variational Autoencoders to model normal, non-defective images, and detecting defects by finding anomalies to this model. In particular, the suggested approach is to utilize recent developments in the learning of “disentangled latent variables” to obtain better representations of the available image data.
Signify will provide the means and methods to obtain the required image data and will take care of a suitable image dataset for developing and testing new visual quality control methods.
Supervision: Loek Tonnaer, Mike Holenderski, Vlado Menkovski
[1] Tonnaer, Loek & Li, Jiapeng & Osin, Vladimir & Holenderski, Mike & Menkovski, Vlado. (2019). Anomaly Detection for Visual Quality Control of 3D-Printed Products. 1-8. 10.1109/IJCNN.2019.8852372.
[2] Doersch, Carl. “Tutorial on variational autoencoders.” arXiv preprint arXiv:1606.05908 (2016).
[3] Locatello, Francesco, et al. “Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations.” International Conference on Machine Learning. 2019.
[4] Mathieu, Emile, et al. “Disentangling Disentanglement in Variational Autoencoders.” International Conference on Machine Learning. 2019.