Certain Investigations on Transform Domain Based Image Denoising Techniques

Abstract

ABSTRACT newlineImage denoising is an important research area in computer vision communities. The newlinemain objective of the research is to develop and implement transform based image newlinedenoising algorithms to remove noise from the images contaminated by Additive White newlineGaussian Noise (AWGN) with preservation of the global contrast, edge structures and newlinetexture information of the image and reduction of blurring and artifacts in the denoised newlineimage. newlineOne of the proposed transform-based denoising techniques is the Hybrid newlineWavelet and Quincunx Diamond Filter Bank (HWQDFB)-based denoising scheme. The newlineHWQDFB combines the Wavelet Filter Bank (WFB) and Quincunx Diamond Filter newlineBank (QDFB) for an efficient representation of images. The QDFB is designed from newlinefinite impulse response filters using Kaiser window with good frequency selectivity and newlinehigh stopband attenuation to reduce aliasing distortion. At first, the HWQDFB newlinedecomposes the noisy image into subbands of different frequencies and orientations newlineusing discrete Meyer wavelet. The QDFB is applied on the detail subbands of wavelet newlinefilter bank to obtain the directional subbands. Then, the denoised detail coefficients are newlinedetermined by the Bayes Least Squares (BLS) estimator from noisy image subband newlinecoefficients modelled as the Gaussian Scale Mixture (GSM). newlineThe HWQDFB-based denoising scheme is experimented with images of newlinediversified characteristics like Lena, Barbara, boat, pepper, circuit and cameraman. It newlinereduces blocking, ringing and staircase artifacts and scratch phenomena in smooth newlineregions with satisfactory visual quality which is measured by visual inspection and newlineperformance measures like Structural Similarity Index Metric (SSIM) and Figure of newlineMerit (FOM). Also, it improves the Peak Signal-to-Noise Ratio (PSNR) with less newlinecomputational complexity. But at high noise densities, this algorithm fails to preserve newlineedges, and fewer artifacts are present in the denoised image. To overcome this newlineslimitation, the Subsampled Pyramid and Nonsubsampled Directional Filter Bank newline(SPNSDFB)-based image