Image and vision systems often require scale transformation of images to different sizes to meet various display requirements. Images may be scaled upwards to fit in a large display, for example enlarging a small image to fit on your desktop wallpaper. Images may also be scaled down when the native resolution is higher than the desired display space. For this project we are particularly interested in up-scaling of images.
The common image up-scaling techniques used today usually result in blurring of the image. Various solutions to this problem have been developed, a common one being the use of isotropic high-pass filters to preserve and/or enhance edges before and/or after scaling. The limitation of this method is that high frequency shot noise in the original image that may be barely perceptible is enhanced by the high-pass filter and then scaled along with the image. The result is a perceptually very noisy scaled version of the original image and noise reduction becomes essential.
Noise reduction in an image is accomplished by the use of a smoothing (low-pass) filter. In addition to smoothing out noise, low-pass filtering also blurs out edges along with the noise. Hence, there is a definite trade-off between how much we can preserve the edges and reduce noise using purely linear filtering. However, with the use of non-linear filters both can be achieved concurrently.
Two common algorithms for edge-preserving smoothing filters are anisotropic
diffusion [4] and bilateral filtering [1]. For this project we focus on the use
of bilateral filters to preserve edges in image up-scaling. We first introduce a
basic image scaling algorithm and then add a linear edge-detection algorithm. We
then explore the limitations of linear edge-preserving filters and
implement the method of bilateral filtering as a solution to these limitations.
Then we compare the performance with and without bilateral filtering. It is
important to note that judging performance of non-linear imaging (and other)
systems is best done qualitatively rather than quantitatively as they tap into
perceptual characteristics of our visual system.
<Previous Page>
<Next Page>