Abstract:
The basic objective of Super Resolution (SR) techniques are to generate a High Resolution
(HR) image from one or more Low Resolution (LR) images. Due to its vast
applications in real world SR is becoming a very active research area and many SR
methods has been proposed in last decade or so. Generally the SR methods can be
divided into three categories Learning based methods, Interpolation based methods and
Reconstruction based methods. In this thesis, a comprehensive study is done to get a
better understanding of benefits and shortcomings of these categories. A position patch
based SR method is also proposed combining the positive effects of all three categories.
The proposed approach take the advantage of non-local self similarity of image to fill the
missing pixels in HR image. Non local similarity uses the redundant frequency information
in image whether it is local or faraway from the observed patch in that image. First
an image is decomposed into many small patches, then the image is searched for similar
patches. After searching the similarity in patches relative to each patch, the HR image
is then reconstructed by filling the missing pixels using information present in similar
patches. The main focus in this method is to preserve the edges in HR reconstructed
image because most methods fails to perform with the same efficiency around edges of
an image compare to their performance in rest of the image. The results show significant
improvement in the quality of image. An enhancement filter is also applied on the
final image to get the best results. A framework for given work is discussed in detail.
The results indicate significant improvement in sharpness and resolution of image. Also
the jagging effect around edges is reduced using proposed technique. The experimental
analysis between proposed and state of the art methods shows the effectiveness of proposed
approach. For quantitative analysis, two techniques PSNR and SSIM are used
to validate the obtained results. For training set, fei database is used. The current
methods fail to perform well when image is different from training sets which is why
different datasets are used for testing purposes.
