EngD student on Deep learning to detect bitmap artefacts

The Electronic Systems (ES) group within the Department of Electrical Engineering of Eindhoven University of Technology (TU/e) and Canon Production Printing are seeking to hire an outstanding EngD candidate within the field of Embedded Computer Vision.

Canon Production printing develops high-end digital inkjet printer systems for the high-volume and graphic arts professional printing markets. In such printers' high-resolution bitmaps are prepared for printing using a fast image processing implementation that is able to match the demanding data rate requirements of the systems. This image processing ensures that the prints are delivered with the best possible print quality and without any artifacts.

To ensure that the image quality targets are met, the image processing implementations are rigorously tested, both using offline tooling and by printing extensive test sets on real printers. However, finding artifacts is a challenge due to the bulk of data, and sometimes feels like looking for a needle in a haystack: Visually inspecting the output bitmaps that contain hundreds of millions of pixels is very labor intensive, as is looking for artifacts in big stacks of paper.

Canon Production Printing is looking at ways to improve the identification of print artifacts in image processing output in order to further improve the quality of our deliveries. One of the planned improvement steps is the integration of machine learning algorithms during regression testing as a way to help the developers identify print artifacts in the image processing output before the software is released to be integrated into the print engine software.

This EngD project will consist of deep learning algorithm development and recreation/simulation of image quality artifacts. To obtain training data for the machine learning, existing (mostly artifact-free) regression test-sets can be used. In addition, miscellaneous types of artifacts can be (re-) created by building past versions of the software, as well as by manipulation of input data to the software. Additional artifacts may be simulated by adjusting the image processing output bitmaps. The training data will be used to create deep neural networks with the goal of flagging test cases where the output is expected to contain artifacts. For this application, a good network should combine high processing speed with good classification performance. The resulting network will be integrated into the build and test environment to run during (daily) regression testing and automated builds, where it should flag the bitmaps most likely to contain image artifacts. These bitmaps can then be inspected further manually.