Acharya, Aditya (2018) Development of Some Spatial-domain Preprocessing and Post-processing Algorithms for Better 2-D Up-scaling. PhD thesis.
|PDF (Full text is restricted up to 30/09/2020)|
Image super-resolution is an area of great interest in recent years and is extensively used in applications like video streaming, multimedia, internet technologies, consumer electronics, display and printing industries. Image super-resolution is a process of increasing the resolution of a given image without losing its integrity. Its most common application is to provide better visual effect after resizing a digital image for display or printing. One of the
methods of improving the image resolution is through the employment of a 2-D interpolation.
An up-scaled image should retain all the image details with very less degree of blurring meant for better visual quality. In literature, many efficient 2-D interpolation
schemes are found that well preserve the image details in the up-scaled images; particularly at the regions with edges and fine details. Nevertheless, these existing interpolation schemes too give blurring effect in the up-scaled images due to the high frequency (HF) degradation during the up-sampling process. Hence, there is a scope to further improve their performance through the incorporation of various spatial domain pre-processing, post-processing and composite algorithms.
Therefore, it is felt that there is sufficient scope to develop various efficient but simple pre-processing, post-processing and composite schemes to effectively restore the HF contents in the up-scaled images for various online and off-line applications. An efficient and widely used Lanczos-3 interpolation is taken for further performance improvement through the incorporation of various proposed algorithms.
The various pre-processing algorithms developed in this thesis are summarized here. The term pre-processing refers to processing the low-resolution input image prior to image
up-scaling. The various pre-processing algorithms proposed in this thesis are: Laplacian of Laplacian based global pre-processing (LLGP) scheme; Hybrid global pre-processing (HGP); Iterative Laplacian of Laplacian based global pre-processing (ILLGP); Unsharp masking based pre-processing (UMP); Iterative unsharp masking (IUM); Error based up-sampling(EU) scheme.
The proposed algorithms: LLGP, HGP and ILLGP are three spatial domain preprocessing algorithms which are based on 4th, 6th and 8th order derivatives to alleviate nonuniform
blurring in up-scaled images. These algorithms are used to obtain the high frequency (HF) extracts from an image by employing higher order derivatives and perform precise
sharpening on a low resolution image to alleviate the blurring in its 2-D up-sampled counterpart.
In case of unsharp masking based pre-processing (UMP) scheme, the blurred version of a low resolution image is used for HF extraction from the original version through image subtraction. The weighted version of the HF extracts are superimposed with the original image to produce a sharpened image prior to image up-scaling to counter blurring effectively.
IUM makes use of many iterations to generate an unsharp mask which contains very high frequency (VHF) components. The VHF extract is the result of signal decomposition in terms of sub-bands using the concept of analysis filter bank. Since the degradation of VHF components is maximum, restoration of such components would produce much better restoration performance.
EU is another pre-processing scheme in which the HF degradation due to image upscaling is extracted and is called prediction error. The prediction error contains the lost high frequency components. When this error is superimposed on the low resolution image prior to image up-sampling, blurring is considerably reduced in the up-scaled images.
Various post-processing algorithms developed in this thesis are summarized in following. The term post-processing refers to processing the high resolution up-scaled image. The various post-processing algorithms proposed in this thesis are: Local adaptive Laplacian (LAL); Fuzzy weighted Laplacian (FWL); Legendre functional link artificial neural network(LFLANN).
LAL is a non-fuzzy, local based scheme. The local regions of an up-scaled image with high variance are sharpened more than the region with moderate or low variance by employing a local adaptive Laplacian kernel. The weights of the LAL kernel are varied as per the normalized local variance so as to provide more degree of HF enhancement to high variance regions than the low variance counterpart to effectively counter the non-uniform blurring. Furthermore, FWL post-processing scheme with a higher degree of non-linearity is
proposed to further improve the performance of LAL. FWL, being a fuzzy based mapping scheme, is highly nonlinear to resolve the blurring problem more effectively than LAL which
employs a linear mapping.
Another LFLANN based post-processing scheme is proposed here to minimize the cost function so as to reduce the blurring in a 2-D up-scaled image. Legendre polynomials are
used for functional expansion of the input pattern-vector and provide high degree of nonlinearity. Therefore, the requirement of multiple layers can be replaced by single layer LFLANN architecture so as to reduce the cost function effectively for better restoration performance. With single layer architecture, it has reduced the computational complexity and hence is suitable for various real-time applications.
There is a scope of further improvement of the stand-alone pre-processing and postprocessing schemes by combining them through composite schemes. Here, two spatial domain composite schemes, CS-I and CS-II are proposed to tackle non-uniform blurring in an up-scaled image. CS-I is developed by combining global iterative Laplacian (GIL) preprocessing scheme with LAL post-processing scheme. Another highly nonlinear composite
scheme, CS-II is proposed which combines ILLGP scheme with a fuzzy weighted Laplacian post-processing scheme for more improved performance than the stand-alone schemes.
Finally, it is observed that the proposed algorithms: ILLGP, IUM, FWL, LFLANN and CS-II are better algorithms in their respective categories for effectively reducing blurring in the up-scaled images.
|Item Type:||Thesis (PhD)|
|Uncontrolled Keywords:||Image up-sampling, image super-resolution, image deblurring, 2-D interpolation, Nearest-neighbour interpolation, Bilinear interpolation, Bicubic interpolation, Lanczos interpolation, DCT interpolation, Laplacian of Laplacian, fuzzy logic, FLANN, fuzzy Laplacian, iterative unsharp masking, low resolution, high resolution|
|Subjects:||Engineering and Technology > Electronics and Communication Engineering > Fuzzy Systems|
Engineering and Technology > Electronics and Communication Engineering > Image Processing
|Divisions:||Engineering and Technology > Department of Electronics and Communication Engineering|
|Deposited By:||IR Staff BPCL|
|Deposited On:||28 Sep 2018 15:51|
|Last Modified:||11 May 2022 11:37|
Repository Staff Only: item control page