Subject

The image restoration problem is key in image processing. It consists of estimating a continuous pristine image given few data samples that are corrupted by optical blur and electronic noise. Using a simple iterative estimation method, parameters of an image representation may be computed given the input data such that the noise part is left aside as a residual. The downsize of iterative estimation technique is the speed of computing and hopefully the Graphics Processing Units (GPUs) may be accelerating by order of magnitudes the computation.

Kind of work

In this Master thesis topic, the student will use the Vulkan API and GLSL shading language that are commonly used for real-time computer rendering in games and interactive visualization. An original image representation model and associated estimation algorithm is currently developed at ETRO and is a perfect fit for such acceleration using parallel graphics rasterization techniques.

Depending on the personal interest of the student, the software will be expanded to address higher-dimensional image restoration problems such as volumetric images in Magnetic Resonance Imaging, or 360 panorama images, or ultra-high definition Gigapixel image processing, 4D light-field, 5D plenoptic reconstructions, etc

The performance gain over a vanilla reference sequential C++ implementation on CPU will be evaluated and the result quality will be compared to alternative techniques, such as deep learning for image upscaling and denoising.

Number of Students

1-2

Expected Student Profile

- The student profile should have good command of C++
- Interest in learning high-performance GPU graphics programming with Vulkan and GLSL as well as learning to use performance profiling tools.