Image processing is in practice in most of the medical, industrial, and military applications. In this process, for accurate analysis of the image(s) by human interpretation or autonomous machine perception, denoising of the image is mandatory. Denoising process helps to obtain original image from the corrupted image. Continuous efforts made by the researchers in this field to improve coding technique or introduce new filtering methods to get betterdenoised images in terms of retaining or recovering original details of the image. Various denoising technique are in use. Improved Non Local Means (NLM) and other denoising techniques proposed in recent years
Quality of denoised image improves by waveletbased denoising technique. Waveletbased image denoising techniques estimate the threshold value by considering either subband or universal thresholding to denoise the image. Visushrink, SURE shrink, Bayes shrink, and Neighsure shrink are some of the thresholding methods to estimate the threshold value. B. K. Shreyamsha Kumar proposed
The paper provides brief introduction of image filters and use of these filters along with different thresholding methods in section 1. Different thresholding techniques discussed in details in section 2. ANLM filter adopted in the present work briefly discussed in section 3. Section 4 deals with methodology followed in this paper through the block diagram along with description of each block. Results and discussions described in the section 5. Section 6 concludes the proposed method.
The universal threshold defined by VisuShrinkis given in (1).
Where
and L  total number of pixels in an image.
However, this technique yields less preserved details since the threshold value is high for large values of L, as it kills signal coefficients and noise.
In the subband adaptive technique proposed by Donoho and Johnstone as Steins Unbiased Risk Estimate (SURE), it leads to select threshold TSURE, which adapts to the data by minimizing the estimation of the mean square error (MSE). Threshold parameter TSURE calculated using (2).
Where,
L – total number of wavelet coefficients of a subband under consideration,
For images corrupted by Gaussian noise, the BayesShrink method is more effective; it uses an adaptive datadriven threshold technique. In this method, soft thresholding and threshold of a sub band are determined by modeling the wavelet coefficients. Random variables are obtained within each subband of an image using Generalized Gaussian Distribution (GGD).
Bayes threshold (T_{B}) of a subband is calculated by (4).
Where
Here
Where
Here
Where
Where, w is modeled as zero mean,
when
This thresholding technique incorporates neighboring wavelet coefficients to estimate the threshold value. The neighborhood window size should be odd. A 3x3 window is depicted in
The shrink function for a particular odd size window is given in (9).
b _{i, j} refers to centered wavelet coefficient.
Where,
(S_{i,j} ) the sum of all the wavelet coefficients within the neighborhood window.
Where,
The + sign indicates to consider only positive values during estimation and becomes zero for negative values.The estimated center wavelet coefficient obtained through a noisy wavelet coefficient using (11).
NeighShrinkSURE is an improved version of NeighShrink in the image denoising technique proposed by Dengwen and Wengang. Stein's Unbiased Risk Estimate (SURE) method used to determine each subband's optimal threshold and neighboring window size (12).
Where, T  threshold,
k  window size and
s  denotes the subband.
Rajiv Verma et al.
The mean gray level of a pixel value in the neighborhood window (W_{k}) of odd number n x n matrix, centered on pixel i is given as (13).
Where, (k, l)  pixel coordinate in neighborhood
The absolute gray level difference
Gray level difference (GLD) of each pixel of an image calculated using (13) and (14).
The Mean
and standard deviation (s) of the GLD is calculated using the formula
Then threshold T_{1} and T_{2} are defined as
where
Then an optimal search window can be made,
Denoised image obtained by adaptive search using the formula (20).
In this process, the noisy image (I) is applied to ANL filter to obtain a prefiltered image (I_{F}). Method noise (O) is obtained by computing the difference of prefiltered image (I_{F}) and noisy image (I) through summing block1. Method noise applied to discrete wavelet transform (DWT) block to obtain three level decomposed images of method noise. NeighShrinksure thresholding eliminates the noisy components of the DWT output. Inverse discrete wavelet transform (IDWT) (I_{W}) recovers the image features present in the method noise. The I_{F} image and image features I_{W} are combined in summing block2 to obtain the denoised image.
The newly developed denoising algorithm implemented through MATLAB software compares with other denoising algorithms. A standard 256 x 256 grayscale images denoised using Wavelet Thresholding (WT), Gaussian Bilateral Filter with Method Noise Thresholding (GBFMT) and Weiner Filter with Residual noise Thresholding (WFRT) methods is compared with proposed denoising algorithm. Standard images of Lena, Barbara, and Girl face considered in the proposed work are shown in
Refers to the ratio of maximum possible power of a signal to the power of corrupting noise.
where, MSE refers to Mean Sample Error of the pixel throughout the image.
The similarity between the original image and the image obtained by the denoising technique is the structural similarity index. It is a perceptionbased model of an image, and change of perception in structural information of an image is referred as image degradation. Image degradation may be due to luminance and contrast masking. Distortions measured at the edges of the denoised image are the luminance masking, and distortions in an image's texture are contrast masking. The structural similarity index is given by (22).
Where
This parameter (IQI) is obtained by modeling the image distortion related to the reference image with loss correlation, luminance distortion, and contrast distortion.
PSNR and SSIN parameters of the denoised image using ANLMNT (proposed method) are compared with WT, GBFMT and WFRT methods.

10 
20 
30 
40 
50 
Method 
Input Image(LENA) 

WT 
27.36 
24.97 
23.97 
23.41 
22.90 
GBFMT 
33.07 
29.23 
27.19 
25.76 
24.69 
WFRT 
33.23 
28.94 
26.71 
25.13 
23.79 
ANLMNT (Proposed Method) 
33.80 
30.34 
28.11 
26.32 
24.75 

Input Image(BARBARA) 

WT 
26.33 
24.27 
23.28 
22.69 
22.29 
GBFMT 
31.78 
28.33 
26.53 
25.27 
24.26 
WFRT 
32.10 
28.31 
26.21 
24.72 
23.48 
ANLMNT (Proposed Method) 
32.54 
29.18 
27.18 
25.59 
24.13 

Input Image(GIRLFACE) 

WT 
30.72 
26.03 
24.80 
23.92 
23.07 
GBFMT 
33.71 
29.59 
27.09 
25.22 
23.84 
WFRT 
32.78 
28.14 
25.66 
23.85 
22.43 
ANLMNT (Proposed Method) 
34.62 
30.62 
27.99 
25.89 
24.20 
The proposed ANLMNT denoising method depicts an improvement in PSNR compared to other denoising techniques. However PSNR value decreases as sigma increases (σ >30). A study with high contrast standard image (Girlface) shows that the proposed work has significantly improved PSNR values compared to other standard images with low contrast (Lena and Barbara). The proposed method also show increase in SSIN values of the standard images (

10 
20 
30 
40 
50 
Method 
Input Image(LENA) 

WT 
0.9989 
0.9985 
0.9983 
0.9982 
0.9981 
GBFMT 
0.9996 
0.9992 
0.9989 
0.9986 
0.9984 
WFRT 
0.9996 
0.9992 
0.9989 
0.9986 
0.9982 
ANLMNT (Proposed Method) 
0.9996 
0.9994 
0.9991 
0.9988 
0.9985 

10 
20 
30 
40 
50 
LENA 
0.9950 
0.9893 
0.9816 
0.9734 
0.9647 
BARBARA 
0.9930 
0.9853 
0.9784 
0.9686 
0.9585 
GIRL FACE 
0.9956 
0.9875 
0.9774 
0.9668 
0.9572 
Noise Filters with 
GBFMT [5] 
WFRT[6] 
FAN [7] 
ANLMNT (Proposed Work) 
PSNR 
29.23 
28.94 
30.46 
30.34 
SSIN 
0.9935 
0.9992 
0.851 
0.9994 
Method Noise Thresholding (MNT) technique implemented through Gaussian Bilateral filter and Wiener Filter with Residual Thresholding shows lesser PSNR and SSIN values when implemented through Adaptive Non Local mean filter.
Comparison
The proposed denoising technique studied for standard Lena image with varying sigma values and wavelets shows improvement in PSNR value for lower sigma values (10, 20) and for higher values of sigma, it is constant. This reveals that proposed ANLMNT technique is more suitable for images with lower sigma values. In addition, PSNR increases with bior6.8 wavelet as compared to other wavelets (
Wavelet 
Db8 
Sym8 
Db16 
Coif5 
Bior6.8 

PSNR Value 

10 
33.75 
33.77 
33.75 
33.77 
33.80 
20 
30.31 
30.32 
30.29 
30.32 
30.34 
30 
28.10 
28.11 
28.06 
28.12 
28.11 
40 
26.32 
26.33 
26.30 
26.31 
26.32 
50 
24.77 
24.77 
24.74 
24.75 
24.75 
An Adaptive NL means of filtering with method noise thresholding (ANLMNT) technique is proposed in this work. Simulation results of the present method depict that the denoised images have improved PSNR and SSIN values when compared with other method noise thresholding techniques. The proposed work is best suited for images with high noise and high contrast. Finetuning at the prefiltering stage may further enhance the quality of the image. Implementation of other denoising techniques using method of noise thresholding may result in improved denoised image. PSNR value substantially reduces for higher noise levels (