|M.Sc Student||Kalit Yaron|
|Subject||On Stochastic Interpolation of Color Textures|
|Department||Department of Electrical and Computer Engineering||Supervisor||ASSOCIATE PROF. Moshe Porat|
Image interpolation is a fundamental and important task that is involved in many image processing operations. Existing image interpolation methods are mostly based on the monochromatic gray-level representation, and are local, i.e. assign values to new pixels according to a function of the neighboring pixels' values. An important class of interpolation algorithms known as super-resolution (SR) algorithms constructs high resolution images from a set of low resolution images. The classical SR techniques combine low resolution images at sub-pixel misalignments. Another family of SR algorithms uses a database of low and high resolution image pairs in order to learn correspondences between low and high resolution patches.
In this work we focus our attention on the interpolation of color textures. Textures constitute an important part of natural images and provide important visual cues for the human visual system. We present an interpolation approach based on a single image that exploits the redundancy and recurring patterns within textures. Our approach combines local information from neighboring pixels with global information based on patch similarity in an adaptive manner that takes into account characteristics of the human visual system.
Motivated by psychophysical studies of texture perception indicating that global image statistics play a major role in the perception of textures, we propose a stochastic super resolution method that is based on statistical relations between pixels of the low resolution textures. We demonstrate that such random choice of values, modified to account for local image characteristics, preserves the appearance of the low resolution textures and is well suited to a wide variety of textures, both stationary as well as non-stationary. Since color images are becoming the default format in most cameras and mobile devices, our method is designed specifically for color images, performing interpolation of both the intensity and color components. Our conclusion is that our method outperforms presently available methods and could be instrumental in designing interpolation systems for color display screens.