Research Of Image Denoising Based On Multi-scale Transform

In the long history of the development of image processing, wavelet is the maindenoising tool, but wavelet analysis can describe the exotic singularity and can not extendedto the line and surface singularitys, so the wavelet transform show many shortcomings in thedenoising. Multiscale geometric analysis technology embodied great advantage in the imagedenoising, it can be more effectively represent high dimensional singular characteristics,which has overcome the defects of wavelet method, so this study is very important and hassignificant meaning in image denoising.This paper focuses on multi-scale transform and thresholding method based onmulti-scale transform. First, the theoretical basis of wavelet transform and multi-scalegeometric transformation are introduced, the nature and advantages or disadvantages of eachtransform from their basic ideas and formulas derivation are summarized, then the noisecharacteristics, the methods of image denoising and evaluation criteria are introduced, focuseson the classical thresholding method; then based on Contourlet transform, nonsubsampledContourlet transform (NSCT) is introduced, and some improved de-noising methods areproposed.Since NSCT has shifted invariance, It is suitable for the stringent requirements of thepositional shift for correlation noising; and correlation retains more complete edgeinformation, different type of coefficients are selected using signal and noise correlation in thetransform domain, the separation of noise is achieved. First, nonsubsampled Contourlettransformation decomposed image noise, using the normalized correlation coefficient methodcombines the adaptive threshold Bayes denoise decomposed coefficients, finally, usinginverse nonsubsampled Contourlet transform with processed coefficients, the image isreconstructed. In this way the advantages of both Bayes threshold value and the effective useof correlation coefficient of NSCT are achieved. The thesis has studied the characteristics ofNSCT’s coefficients, different sub-band should be required to choose different threshold, theenergy of sub-band coefficients is approximately equal to the square of the NSTC coefficients,the energy is proportional to the contour information contained in the image, so we can usethe energy to adjust the threshold adaptively. Containing energy coefficient ratio of Bayesadaptive threshold for image decomposition’s NSCT coefficients hard thresholding, thismethod can adaptively adjust the threshold value. Experiments show that the improvedmethod can effectively remove noise and Gibbs artifacts, resulting better simulation visualeffects and higher Peak-value Singal to Noise Ratio(PSNR).Wavelet denoising threshold rule can also be applied to Curvelet domain. Because of theshortcomings of hard threshold, we add the adjustment factor in the threshold estimator,getting soft and hard threshold compromise function. First appropriately and effectivelycorrected the scale parameter of NormalShrink threshold, an improved threshold value is obtained. Utilization soft and hard threshold compromise function process the high-frequencycoefficients of after Curvelet transform decomposition after decomposition. Finally, using theprocessed Curvelet coefficient, the image is reconstructed, resulting in the denoised image.Simulation results show that, the improved Curvelet denoising has retained the contourfeatures of the image very well, and improved image quality and increased the PSNR.