Page 1
Signal and Image Denoising smaller than threshold, set to zero; otherwise it is kept or modified. Replacing the small noisy coefficients by zero and inverse wavelet transform on the result may lead to reconstruction with the essential signal characteristics and with less noise. But, wavelet transform suffers due to poor directionality and does not provide a geometrically oriented decomposition in multiple directions.
Page 2
Page 3
Signal and Image Denoising threshold value. The proposed threshold mentioned in this paper shows better performance (PSNR = 47.2) over other techniques. (iii) Wavelet Transform Based Image Denoise Using Threshold Approaches by Akhilesh Bijalwan et al. (Jun 2012) shown a simple and sub band semi soft threshold proposed method to address the issue of image recovery from its noisy counterpart. It is based on the discrete wavelet transform and Gaussian distribution modeling of sub band coefficients. The image denoise algorithm uses semi thresholding to provide smoothness and better image details preservation. The wavelet semi soft thresholding denoise algorithm produce overall better PSNR and MSE result compared with other traditional denoise approaches. (iv) Image Denoising Using Wavelet Threshold Methods by Namrata Dewangan et al. (Jun 2012) shown that Bivariate shrinkage function is used to reduce Gaussian, Salt & Pepper and speckle noise at different resolution levels. These results obtained have shown significant noise reduction then standard denoising methods such as Sure Shrink, Bayes Shrink, Neigh Shrink and Block Shrink. The focus was shifted from the Spatial and Fourier domain to the Wavelet transform domain. Ever since Donohos Wavelet based thresholding approach was published in 1995, there was a surge in the denoising papers being published. Although Donohos concept was not revolutionary, his methods did not require tracking or correlation of the wavelet maxima and minima across the different scales as proposed by Mallat [1]. Thus, there was a renewed interest in wavelet based denoising techniques since Donoho [2] demonstrated a simple approach to a difficult problem. Researchers published different ways to compute the parameters for the thresholding of wavelet coefficients. Data adaptive thresholds [3] were introduced to achieve optimum value of threshold. Later efforts found that substantial improvements in perceptual quality could be obtained by translation invariant methods based on thresholding of an Undecimated Wavelet Transform [4]. These thresholding techniques were applied to the nonorthogonal wavelet coefficients to reduce artifacts. Multiwavelets were also used to achieve similar results. Probabilistic models using the statistical properties of the wavelet coefficient seemed to outperform the thresholding techniques and gained ground. Recently, much
E&CE Dept GNDECB
Page 4
Signal and Image Denoising effort has been devoted to Bayesian denoising in Wavelet domain. Hidden Markov Models and Gaussian Scale Mixtures have also become popular and more research continues to be published. Tree Structures ordering the wavelet coefficients based on their magnitude, scale and spatial location have been researched. Data adaptive transforms such as Independent Component Analysis (ICA) have been explored for sparse shrinkage. The trend continues to focus on using different statistical models to model the statistical properties of the wavelet coefficients and its neighbors. Future trend will be towards finding more accurate probabilistic models for the distribution of nonorthogonal wavelet coefficients.
Page 5
where = (y). It should be clear that in practice the function f is unknown, so the MSE is usually estimated. Note that relation (3.1) is not genetal, since the noise may be non-additive and the relation between the observed signal and the original signal may be stochastic. Nevertheless, (3.1) is a good model for many practical situations. It is assumed in what follows, without loss of generality, that ti are within the unit interval [0,1]. Furthermore, for simplicity, it is assumed that these sample points are equally spaced and that n = 2^J for some J N.
Page 6
Page 7
Signal and Image Denoising or kill rule. The different types of thresholding techniques in wavelet domain are discussed here.
This work is focused on the wavelet transform filtering method. All wavelet transform denoising algorithms involve the following three steps in general Forward Wavelet Transform: Wavelet coefficients are obtained by applying the wavelet transform. Estimation: Clean coefficients are estimated from the noisy ones. Inverse Wavelet Transform: A clean image is obtained by applying the inverse wavelet transform. There are many ways to perform the estimation step. Broadly, they can be classified as: Thresholding methods Shrinkage methods Other approaches 4.1 Thresholding Methods
Page 8
Signal and Image Denoising These methods use a threshold and determine the clean wavelet coefficients based on this threshold. There are two main ways of thresholding the wavelet coefficients, namely the hard thresholding method and the soft thresholding method. 4.1.1 Hard and soft thresholding Two standard thresholding policies are: hard-thresholding, (keep or kill), and soft-thresholding (shrink or kill). In both cases, the coefficients that are below a certain threshold are set to zero. In hard thresholding, the remaining coefficients are left unchanged Thard(w) = 0, if |w| T, w, if |w| > T. In soft thresholding, the magnitudes of the coefficients above threshold are reduced by an amount equal to the value of the threshold. Tsoft(w) = 0, if |w| T, sgn(w)(|w| T), if |w| > T. We can say that in both cases each wavelet coefficient is multiplied by a given shrinkage factor, which is a function of the magnitude of the coefficient. In soft thresholding, the estimates are biased: large coefficients are always reduced in magnitude; therefore, the mathematical expectations of their estimates differ from the observed values. The reconstructed image is often over smoothed. On the other side, a disadvantage of the hard thresholding is its abrupt discontinuity: estimates have a larger variance and may be highly sensitive to small changes in the data. In practice, especially when the noise level is high, hard thresholding yields abrupt artifacts in the reconstructed image. Due to this, in image processing applications the soft thresholding is usually preferred over the hard one. Various thresholding policies have been proposed, which are a compromise between the classical hard and soft ones, like, e.g., the hyperbola function Thyper(w) = sgn(w)w2- T2 for |w| > T, and Thyper(w) = 0 otherwise; it attenuates large coefficients less than soft thresholding, and is continuous.
E&CE Dept GNDECB
Page 9
Signal and Image Denoising Soft thresholding is used for all the algorithms due to the following reasons: Soft thresholding has been shown to achieve near minimax rate over a large number of Besov spaces [2]. Moreover, it is also found to yield visually more pleasing images. Hard thresholding is found to introduce artifacts in the recovered images. We now study three thresholding techniques- VisuShrink, SureShrink and BayesShrink. 4.1.2 VisuShrink VisuShrink is thresholding by applying the Universal threshold proposed by Donoho and Johnstone [5]. This threshold is given by 2log M where is the noise variance and M is the number of pixels in the image. It is proved in [5] that the maximum of any M values iid as N(0, 2 ) will be smaller than the universal threshold with high probability, with the probability approaching 1 as M increases. Thus, with high probability, a pure noise signal is estimated as being identically zero. However, for denoising images, VisuShrink is found to yield an overly smoothed estimate. This is because the universal threshold (UT) is derived under the constraint that with high probability, the estimate should be at least as smooth as the signal. So the UT tends to be high for large values of M, killing many signal coefficients along with the noise. Thus, the threshold does not adapt well to discontinuities in the signal. 4.1.3 SureShrink SureShrink is a thresholding by applying subband adaptive threshold, a separate threshold is computed for each detail subband based upon SURE (Steins unbiased estimator for risk), a method for estimating the loss || u|| 2 in an unbiased fashion. The results obtained for the image on my picture using SureShrink are much better than VisuShrink. The sharp features of image are retained and the MSE is considerably lower. This because SureShrink is subband adaptive. 4.1.4 BayesShrink BayesShrink [6] is an adaptive data-driven threshold for image denoising via wavelet soft-thresholding. The threshold is driven in a Bayesian framework, and we
E&CE Dept GNDECB
Page 10
Signal and Image Denoising assume generalized Gaussian distribution (GGD) for the wavelet coefficients in each detail subband and try to find the threshold T which minimizes the Bayesian Risk.. We found that BayesShrink performs better than SureShrink in terms of MSE. The reconstruction using BayesShrink is smoother and more visually appealing than one obtained using SureShrink. 4.2 Shrinkage Methods These methods shrink the wavelet coefficients as follows =. x where 0 <= <= 1 is the shrinkage factor. The following methods belong to this category: 4.2.1 Linear MMSE Estimator Michak proposed the linear Minimum Mean Square Estimation (MMSE) method using a locally estimated variance. Two methods were presented for the estimation of the local variance x,k. The first one uses an approximate maximum likelihood (ML) estimator and the second approach uses the maximum a posteriori (MAP). Michak showed that the MAP estimator produces better results compared to the ML estimator. It is suggested that it can be set to the inverse of the standard deviation of the wavelet coefficients that were initially denoised using the ML estimator. 4.2.2 Bivariate Shrinkage using Level Dependency All the above algorithms use a marginal probabilistic model for the wavelet coefficients. However, the wavelet coefficients of natural images exhibit high dependency across scale. For example, there exists a high probability of a large child coefficient if the parent coefficient is large. Sunder and Selesnick in [7] proposed a bivariate shrinkage function using the MAP estimator and the statistical dependency between a wavelet coefficient and its parent. 4.3 Other Approaches 4.3.1 Gaussian Scale Mixtures
Page 11
Signal and Image Denoising Portilla et al. [8] proposed a method for removing noise from digital images based on a statistical model of the coefficients of an over-complete multi-scale oriented basis. Neighborhoods of coefficients at adjacent positions and scales are modeled as the product of two independent random variables: a Gaussian vector and a hidden positive scalar multiplier. The latter modulates the local variance of the coefficients in the neighborhood, and is able to account for the empirically observed correlation between the coefficients amplitudes. Mathematically, the denoising problem can be written as Y = zU +W Where U is the zero mean Gaussian random variable, z is the positive scaler multiplier, W is the AWGN and Y refers to the observed coefficients in the neighborhood. 4.3.2 Non-Local Mean Algorithm Natural images often have a particular repeated pattern. Baudes et al. [8] used the self-similarities of image structures for denoising. As per their algorithm, a reconstructed pixel is the weighted average of all the pixels in a search window. The search window can be as large as the whole image or even span multiple images in a video sequence. Weights are assigned to pixels on the basis of their similarity with the pixel being reconstructed. While assessing the similarity, the concerned pixel, as well as its neighborhood is taken into consideration. This algorithm is computationally intensive. A faster implementation with improved computation performance was later presented by Wang et al. [9]. 4.4 Image Denoising using Thresholding The problem boils down to finding an optimal threshold such that the mean squared error between the signal and its estimate is minimized. The wavelet decomposition of an image is done as follows: In the first level of decomposition, the image is split into 4 subbands, namely the HH, HL, LH and LL subbands. The HH subband gives the diagonal details of the image; the HL subband gives the horizontal features while the LH subband represents the vertical structures. The LL subband is the low resolution residual consisting of low frequency components and it is this subband which is further split at higher levels of
E&CE Dept GNDECB
Page 12
Signal and Image Denoising decomposition. The different methods for denoising we investigate differ only in the selection of the threshold. The basic procedure remains the same: 1. Calculate the DWT of the image. 2. Threshold the wavelet coefficients. (Threshold may be universal or subband adaptive) 3. Compute the IDWT to get the denoised estimate. 4.5 Introduction to wavelet 4.5.1 The wavelet concept and its origins The central idea to wavelets is to analyze (a signal) according to scale. Imagine a function that oscillates like a wave in a limited portion of time or space and vanishes outside of it. The wavelets are such functions: wave-like but localized. One chooses a particular wavelet, stretches it (to meet a given scale) and shifts it, while looking into its correlations with the analyzed signal. This analysis is similar to observing the displayed signal (e.g., printed or shown on the screen) from various distances. The signal correlations with wavelets stretched to large scales reveal gross (rude) features, while at small scales fine signal structures are discovered. It is therefore often said that the wavelet analysis is to see both the forest and the trees. In such a scanning through a signal, the scale and the position can vary continuously or in discrete steps. The latter case is of practical interest in this thesis. From an engineering point of view, the discrete wavelet analysis is a two channel digital filter bank (composed of the lowpass and the highpass filters), iterated on the lowpass output. The lowpass filtering yields an approximation of a signal (at a given scale), while the highpass (more precisely, bandpass) filtering yields the details that constitute the difference between the two successive approximations. A family of wavelets is then associated with the bandpass, and a family of scaling functions with the lowpass filters. 4.5.2 What makes wavelets useful in signal processing? In signal processing, the representation of signals plays a fundamental role. For example, the Arabic numeral representation permits one to easily notice a power of 10,
E&CE Dept GNDECB
Page 13
Signal and Image Denoising but more difficult to realize a power of 2. With the binary representation, the situation is reversed. Meyer wrote any particular representation makes certain information explicit at the expense of information that is pushed into the background and may be quite hard to recover. The Fourier representation reveals the spectral content of a signal, but makes it impossible to recover the particular moment in time (or the particular space coordinates in case of images) where a certain change has occurred. This makes the Fourier representation inadequate when it comes to analyzing transient signals. In signal and image processing, concentrating on transients (like, e.g., image discontinuities) is a strategy for selecting the most essential information from often an overwhelming amount of data. In order to facilitate the analysis of transient signals, i.e., to localize both the frequency and the time information in a signal, numerous transforms and bases have been proposed. Among those, in signal processing the wavelet and the short time Fourier Transform (STFT) are quite standard. Other main advantageous properties of the wavelet transform are: Multiresolution - a scale invariant representation; Edge detection - large wavelet coefficients correspond to image edges; Sparsity - the wavelet transform of natural images tends to be sparse; Fast algorithms - the complexity of the fast discrete wavelet transform is a linear function of the number of the input samples. Besides, the wavelet analysis has been also often motivated from the viewpoint of psychophysical aspects of human vision: it corresponds well to the way how we perceive images. 4.6 Wavelet Transform A wave is usually defined as an oscillating function in time or space. Sinusoids are an example. Fourier analysis is a wave analysis. A wavelet is a small wave that has its energy concentrated in time and frequency. It provides a tool for the analysis of transient, non-stationary, and time-varying phenomena. It allows simultaneous time and frequency analysis with a flexible mathematical foundation while retaining the oscillating wave-like
E&CE Dept GNDECB
Page 14
Signal and Image Denoising characteristic. Figure 4.2 shows the difference between a sine wave and a wavelet. A simple high level introduction to wavelets can be found in the articles by Daubechies.
Fig 4.2
A signal or a function f(t) can often be better analyzed if it is expanded as f(t) = k Cj0,k j0,k(t) + k j>j0 dj,k j,k(t) where both j and k are integer indices. j,k(t) represents the wavelet expansion functions, and j,k(t) represents the scaling functions. They usually form an orthogonal basis. This expansion is termed as wavelet expansion. The term related to the scaling coefficients captures the average or coarse representation of the signal at the scale j0. The term related to the wavelet coefficients captures the details in the signal from scale j0 onwards. The set of expansion coefficients (cj0,k and dj,k) is called the discrete wavelet transform (DWT) of f(t). The above expansion is termed as the inverse transform. Multi resolution analysis (MRA) and Quadrature mirror filters (QMF) are also important for evaluating the wavelet decomposition. In multi resolution formulation, a single event is
E&CE Dept GNDECB
Page 15
Signal and Image Denoising decomposed into fine details [8]. A quadrature mirror filter consists of two filters. One gives the average (low pass filter), while the other gives details (high pass filter). These filters are related to each other in such a way as to be able to perfectly reconstruct a signal from the decomposed components [3]. Three levels of multi resolution analysis and synthesis are shown in Figure 1.6. QMF filters achieve perfect reconstruction of the original signal. Decimation operations are not shown in Figure 4.3.
Page 16
Signal and Image Denoising Decimation operations when removed result in more data samples in multi resolution domain. This redundancy helps in denoising. The two dimensional (2D) wavelet transform is an extension of the one dimensional (1D) wavelet transform. To obtain a 2D transform, the 1D transform is first applied across all the rows and then across all the columns at each decomposition level. Four sets of coefficients are generated at each decomposition level: LL as the average, LH as the details across the horizontal direction, HL as the details across the vertical direction and HH as the details across the diagonal direction. There are other flavors of the wavelet transform such as translation invariant, complex wavelet transform etc., which give better denoising results. The translation invariant wavelet transform (TIWT) performs multi resolution analysis by filtering the shifted coefficients as well as the original ones at each decomposition level. TIWT is shift invariant (also known as time invariant). This approach produces additional wavelet coefficients (possessing different properties) from the same source. This redundancy improves the denoising performance. Complex wavelet transforms (CWTs) are a comparatively recent addition to the field of wavelets. A complex number includes some properties that can not be represented by a real number. These properties provide better shift-invariant feature and directional selectivity. However, CWTs with perfect reconstruction and good filter properties are difficult to develop. Dual tree complex wavelets (DT CWTs) were proposed by Kingsbury [10]. DT CWTs have some good properties such as reduced shift sensitivity, good directionality, perfect reconstruction using linear phase filters, explicit phase information, fixed redundancy and effective computation in O(N). Multi wavelets are wavelets generated by more than one scaling function, while scalar wavelets use only one scaling function. Multi wavelets also improve denoising performance as compared to the scalar wavelet [11]. Wavelet transforms which generate more wavelet coefficients than the size of the input data are termed redundant or over complete. This added redundancy improves the denoising performance. 4.6.1 The Wavelet Transform in Image Denoising
E&CE Dept GNDECB
Page 17
Signal and Image Denoising The goal of image denoising is to remove noise by differentiating it from the signal. The wavelet transforms energy compactness helps greatly in denoising. Energy compactness refers to the fact that most of the signal energy is contained in a few large wavelet coefficients, whereas a small portion of the energy is spread across a large number of small wavelet coefficients. These coefficients represent details as well as high frequency noise in the image. By appropriately thresholding these wavelet coefficients, image denoising is achieved while preserving fine structures in the image. The other properties of the wavelet transform that help in the image denoising are sparseness, clustering, and correlation between neighboring wavelet coefficients [2]. The wavelet coefficients of natural images are sparse. The histogram of the wavelet coefficients of natural images tends to peak at zero. As they move away from zero, the graph falls sharply. Wavelet coefficients also tend to occur in clusters. They have very high correlation with the neighboring coefficients across scale and orientation. All these properties help in differentiating the noise from the signal and enabling its optimal removal. As Burrus and others [3] have concluded, The size of the wavelet expansion coefficients aj,k or dj,k drop off rapidly with j and k for a large class of signals. This property is called being an unconditional basis and it is why wavelets are so effective in signal and image compression, denoising, and detection. Here aj,k are average coefficients, dj,k are detailed coefficients, j are scale indices, and k are translation indices. Donoho shows that wavelets are near optimal for compression, denoising, and detection of a wide class of signals. The following shows the original image with noise and denoised image.
Page 18
Signal and Image Denoising Following are the results of running the denoising algorithms using the soft thresholding method. The denoising is done after adding the Gaussian noise. The reconstructed image is often over smoothed. On the other side, a disadvantage of the hard thresholding is its abrupt discontinuity: estimates have a larger variance and may be highly sensitive to small changes in the data. In practice, especially when the noise level is high, hard thresholding yields abrupt artifacts in the reconstructed image. Due to this, in image processing applications the soft thresholding is usually preferred over the hard one.
Page 19
Page 20
Signal and Image Denoising This work can be further enhanced to denoise the other types if images.it will provide a good add on the already existing denoising techniques used for denoising these images. Moreover, for future work we can use same algorithm for various techniques like fuzzy logic or neural network in order to attain the best output without performing calculations for each and every combination.
REFERENCES
[1] David L. Donoho and Iain M. Johnstone. Adapting to Unknown Smoothness via Wavelet Shrinkage. [2] C. Sidney Burrus, Ramesh A. Gopinath, and Haitao Guo. Introduction to Wavelets and Wavelet Transforms A Primer. [3] David L. Donoho. Unconditional bases are optimal bases for data compression and for statistical estimation. Applied and Computational Harmonic Analysis, [4] Dane Mackenzie. WAVELETS: Seeing The Forest and the Trees [5] Wikipedia, Image noise, Internet: http://en.wikipedia.org/wiki/Image noise [Jan.18, 2009]. [6] David L. Donoho. De-noising by soft-thresholding. [7] Levent Sundur, and Ivan W. Selesnick. Bivariate Shrinkage Functions for WaveletBased Denoising Exploiting Interscale Dependency. [8] A. Buades, B. Coll, and J. M. Morel. A review of image denoising algorithms, with a new one.
E&CE Dept GNDECB
Page 21
Signal and Image Denoising [9] Jin Wang, Yanwen Guo, Yiting Ying, Yanli Liu, and Qunsheng Peng. Fast non-local Algorithm for Image Denoising. [10] N. G. Kingsbury. Complex wavelets for shift invariant analysis and filtering of signals. [11] Dongwook Cho. Image Denoising Using Wavelet Transforms [12] Javier Portilla, Vasily Strela, Martin J. Wainwright, and Eero P. Simoncelli. Image Denoising Using Scale Mixtures of Gaussian in the Wavelet Domain.
APPENDIX
MATLAB CODE
%% De-Noising Signals and Images % The general de-noising procedure involves three steps. The basic version of % the procedure follows the steps described below: % % * Decompose: Choose a wavelet, choose a level N. Compute the wavelet % decomposition of the signal at level N. % % * Threshold detail coefficients: For each level from 1 to N, select a % threshold and apply soft thresholding to the detail coefficients. % % * Reconstruct: Compute wavelet reconstruction using the original % approximation coefficients of level N and the modified detail % coefficients of levels from 1 to N. % % Two points must be addressed in particular: % % * how to choose the threshold, %
Page 22
Page 23
Page 24
Page 25
Page 26