Image Denoising And Fusion Based On Local Features In Wavelet Domain

This thesis is aimed at the analysis and development of local image features in thespatial and transform domain and their applications in image denoising and imagefusion using a wavelet transform. All works that have been done are supposed to dealwith several problems related to the detail blurring in the denoised images caused bycoefficients alignment across scales of wavelet transform, deficiency of the brute-forcesearching from bottom to top for the desired regular image geometry throughout aBandlet transform, the bleaching effect seen in the enhanced low light visual image by aexponential function of a registered infrared image, the discussion of those availabledefinitions of contrast in a wavelet domain and their influence on the performance ofimage fusion. The solutions to these problems are presented here that include concepts,definitions and relative algorithms. Experimental results are also shown to verify thegood performance of the proposed algorithms and to compare them to those results fromother relative methods. The main contributions can be summarized as follows:1. Being image quality indexes, both UIQI and SSIM miss the local backgroundlightness. This will lead to deficiency when these two indexes are used to computeobjective quality measures of images, as they will violate from the perception of humanvision systems concerning the real contrast. To fix this, a new universal image qualityindex is proposed, with emphasis on the agreement of subjective and objectiveevaluation and the HVS sensitivity of contrast changes, i.e., the perception of details.2. In the original second generation Bandlet transform, a brute-force search iscarried out from bottom to top for regular image geometries, which leads to someunnecessary computations. In the light of the relation between the total variation and thelength of object borders, a novel search strategy from top to bottom is devised to avoidunnecessary image partition and searching for geometries. The key idea lies in the valueof the total variation of an area; in the case of zero total variation, there is no geometryat all, so no segmentation and geometry search is needed; as for those homogeneousareas, no benefit will be added through further partition of an area. In this way, the timecomplexity can be decreased since the wavelet coefficients are sparse expression ofimages, these homogeneous areas and zero areas can be guaranteed.3. There are several statistical models used when noisy images are denoised withwavelet methods, such as Marko chains, Markov trees and mixed Gaussian distributionare typical stochastic processes to express the coefficients relation between neighbourscales and among the same scale. By contrast, an interesting concrete method can predict coefficients across scales without any respect of the statistical distribution ofwavelet coefficients. A threshold function based on SURE principle can be devised andinserted into a linear systems to fulfill the denoising by simply solve the linear systems.However, there are two problems exist. Firstly, a coefficient with big value in a finescale may disappear in a coarse scale because of the sub-sampling. Secondly, acoefficient of big value in a coarse scale may not be corresponding to a coefficient ofbig value in a fine scale, and vice versa. In either case, the pixel-wise prediction willlead to blur image features after thresholding and denoising. Inspired by the directionalgeometric flow used in a Bandlet transform, we propose to predict coefficients acrossscales based on the directional geometric flow, which is assumed scale-invariant afterthe thresholding process. This flow-wise prediction will fix the missing coefficient andblur texture details by interpolation along geometry directions.4. An intensity transformation function of infrared images is presented and used forcontext enhancement of visual images, upon which a new image fusion method in theshift-invariant wavelet domain is developed. The function behaves like a sigmoidfunction and shifts and expands the range of dark pixels of infrared images. Theseadjustments according to the local histogram characteristics can avoid artificial brightpixels introduced in the later enhancing of visual images and the bleaching effect in thefinal fused images owing to the exponential map of very dark pixels of the infraredimages. The key idea is the fact that the bleaching effect are owning to the very darkpixels of the infrared image through the used exponential function.5. It is a hot research area that fusing images by selecting coefficients according toa certain contrast (saliency measure) through a multiresolution analysis. However, mostknown definitions of contrast (saliency measure) miss the importance of localbackground lightness, or compute the contrast without considering a sliding window ofproper size. In fact, even the approximate coefficients need a contrast to have a betterfusion results instead of simply averaging source ones. These usually lead to decreasedlightness, prevailing of information from infrared image, and loss of details in theresulted images. By comparison and analysis of those available contrast (saliencymeasure) definitions, a universal contrast is devised and used to develop a new fusionalgorithm to allow the feature from source images enter into the fused image with moresharper contrast. Experimental results verify its fusion performance and its stability.