Abstract:
Super-resolution is a technique of producing a high-resolution (HR) image from one or
more low-resolution (LR) images. Classical interpolation based magnification techniques
like nearest-neighbor, bilinear and bicubic interpolation results in a larger image along
with undesirable artifacts like blurring, aliasing and ringing effects. So the aim of superresolution
is to provide a larger image with good quality (quality means an image with less
undesirable artifacts). Previous super-resolution techniques are based on using multiple
images and learning based methods but the idea here is to use a single image in the superresolution
process.
In this thesis combination of wavelet transform and interpolation based technique to
achieve the super-resolution of a single image. First the edges of the image are boosted
using wavelet transform. After boosting the edges the resultant image undergoes the
process of magnification which is achieved using an interpolation based method.
Interpolation based magnification algorithm produces high-resolution images that is free of
undesirable artifacts.
A comparison of this algorithm with other techniques proposed by other authors is also
done to provide the quantitative and qualitative result to prove the effectiveness of the
methods. For this purpose 85 test images were taken that belonged to different image
categories, on which different super-resolution techniques were applied to access the
effectiveness of the algorithm. In quantitative analysis, the quality measures used are
correlation coefficient, mean-squared error and peak signal-to-noise ratio. The values of
these measures suggested that the proposed algorithm produces good results.
One of the conclusion derived during the analysis of algorithms results is that proposed
algorithm cannot be applied to all categories of images. This technique is successful on
those images that has less edges e.g. for crowd and satellite images, this technique of
super-resolution is not appropriate. But this technique produces superior results for those
images that have fewer edges e.g. face and object images.
