Student Projects WS 2023/24 (ZEuS link)
Work Group Multimedia Signal Processing
Prof. Dr. Dietmar Saupe
Room Z709, email firstname.lastname@example.org
Registration deadline in ZeUS: Oct 15 - Nov 15, 2023
Due date projects: Jan 15, 2024
Next opportunity to register after the deadline: Jan 15 - Feb 15, 2024, with projects due on April 15
- When interested in one of the projects, contact us asap (email@example.com), and we will invite you for a meeting to talk and discuss the project/thesis prospects.
- These projects and the following theses can be scaled such that they are suitable for either Bachelor or Master level.
- We do not offer accompanying seminars. Instead, we arrange and advise a topic related to a chosen project to be presented in an existing seminar in another research group.
There is also an 7 for students to engage in the form of a Student Assistant job (HiWi).
Available projects in Visual Quality Assessment
Project: Intrinsic image resolution
Project: Adaptive psychophysical methods to assess image quality
Project: Robustness of outlier detection against adversarial attacks
Project: Gamification for perceptual quality assessment
Project: Perceptual Quality Assessment of 3D Point Clouds using JND methodology
Project: Multidimensional color perception and image quality assessment
Project: Intrinsic image resolution
Advisor: Prof. Dietmar Saupe
Consumer digital camera systems such as smartphones record images at very high resolutions. 45, 60, and even 100 megapixels are within reach at lower prices than ever, and the trend isn’t really slowing down. However, for larger resolutions, the sensor area per pixel is lower, which increases sensor noise and reduces image quality. For example, the left image is a crop of a smartphone image of resolution 3968x2976, and it shows pixelation artifacts and graininess.
However, an image, derived from a downscaled version (i.e. lower resolution), then upscaled to the same size, does not have such severe artifacts.
The visual quality of the same image, derived from lower resolution, is better!
If we reduce the resolution even further, details will be lost and eventually the crop will show as a uniformly colored flat area.
Thus, there is a sweet spot, and at that certain resolution, the visual quality is best. This is what we call the intrinsic image resolution. It describes the perceivable image resolution more faithfully than the logical pixel resolution. It is the goal of this project to set up and carry out experiments to measure the intrinsic camera resolutions.
This topic is to be differentiated from that of intrinsic camera resolution. There are numerous techniques to define and measure “effective”, “intrinsic”, or “true” camera and lens resolution (see an example below). Here we are interested in perceived digital image resolution, regardless of their source and their processing after capture.
- Familiarization with image filtering and downscaling from reading assignments (text books) with hands-on experience using, e.g., Matlab as a programming environment.
- Implement at least two different rendering methods for images at different resolution. One can be based on low pass image filtering with the cutoff frequency as a parameter.
- Create an interactive viewer: Load e.g. 50 or 100 images of one source with differing resolutions into memory. A slider below the image controls the resolution, image is displayed with zooming and panning enabled.
- Design a lab-based experimental investigation of intrinsic image resolution. Collect suitable raw images, create and test the user interface. Should we allow users to zoom and pan or should we just display a center crop?
- Carry out an experiment to assess the intrinsic resolution for the dataset from the previous step, analyse the results and produce corresponding tables and visualizations.
- Write the project report and give a final presentation.
- For the thesis:
- A large-scale database will be created using crowdsourcing.
- Based on the dataset, an objective (algorithmic) method shall be developed to predict intrinsic image resolution. This can be done with machine learning using handcrafted features or by deep learning with neural networks.
Good programming skills. Interest in signal and image processing. Previous course credit in such a class helpful, but not required.