Datorseende (CVL)

The scientific goal of this project is to develop new methods for tensor-based variational and PDE-based image processing. Methods based on tensor-controlled evolution or diffusion processes are non-linear, iterative approaches to image enhancement that achieve state-of-the-art results. Controlling the filtering process by tensors allows for an adaptive smoothing of image contents, i.e., structures are enhanced and noise is suppressed.

Adaptive image enhancement has many applications, for instance denoising low-dose images (low illumination or medical images), up- and resampling of images (super-resolution and anti-aliasing), pre-processing for image sequence compression (flat regions contain less information), and image arts (painterly rendering). Methods developed for image en- hancement are also applied in estimation problems, as optical flow extraction for motion estimation and disparity estimation from stereo images.

The proposed work aims at extending existing work by introducing new types of tensor control. We intend to consider three aspects of tensor-based diffusion:

• extending the class of variational formulations to tensor-based iterative filtering
• a systematic extension of control tensor computation based on natural image statistics and state-of-the-art color space models

• a systematic extension of tensor based diffusion using the gradient energy tensor and its natural image statistics