 Research
 Open Access
 Published:
Improved BM3D image denoising using SSIMoptimized Wiener filter
EURASIP Journal on Image and Video Processing volumeÂ 2018, ArticleÂ number:Â 25 (2018)
Abstract
Image denoising is considered a salient preprocessing step in sophisticated imaging applications. Over the decades, numerous studies have been conducted in denoising. Recently proposed Block matching and 3D (BM3D) filtering added a new dimension to the study of denoising. BM3D is the current stateoftheart of denoising and is capable of achieving better denoising as compared to any other existing method. However, there is room to improve BM3D to achieve highquality denoising. In this study, to improve BM3D, we first attempted to improve the Wiener filter (the core of BM3D) by maximizing the structural similarity (SSIM) between the true and the estimated image, instead of minimizing the mean square error (MSE) between them. Moreover, for the DConly BM3D profile, we introduced a 3D zigzag thresholding. Experimental results demonstrate that regardless of the type of the image, our proposed method achieves better denoising performance than that of BM3D.
1 Introduction
There are different types of noise that can contaminate a digital image. Depending on the noise type, there are various algorithms present in the literature for denoising the image. Block matching and 3D (BM3D) filtering [5] is one such popular algorithm that reduces additive white Gaussian noise (AWGN) [16] from digital images. In terms of denoising performance, BM3D is considered the best denoising filter to date. It exhibits remarkable results when compared to other existing methods. BM3D works in two identical steps. In the first step, it generates a basic estimate of the noisy image using hard thresholding. Then in the second step, it uses Wiener filter to actually denoise the noisy image. To do so, BM3D uses the basic estimate generated from the first step as an oracle (i.e., a pilot signal) in the Wiener filter.
Wiener filter is an ageold benchmark for image denoising and restoration [23]. This filter needs a degradation function for denoising or restoration. The better the degradation function is, the more denoising is achievable by Wiener filter. BM3D uses the basic estimated image from the first step as the degradation function of Wiener filter. Thus, the ultimate performance of BM3D largely depends on how good the basic estimate is.
Although BM3D achieves good denoising performance, it is not sufficient to denoise images contaminated by huge levels of noise. In other words, the performance of BM3D decreases with the increase of noise level. Again, among the different profiles of BM3D (a profile is a specific set of parameters), the DConly profile (meaning that the 3D transform used is the 3DDCT) generally performs poorer than the others. Therefore, there is scope to either propose better denoising technique than BM3D or to make BM3D perform better than what it can currently do.
The Wiener filter [23] was proposed about half a century ago. Different researchers attempted to improve the performance of the Wiener filter; however, most studies did not directly address one persisting problem of the Wiener filter which is it uses an objective function, called mean square error (MSE), which is often a misleading measure. In other words, it is possible to use a better measure than MSE as the objective function of Wiener filter. Also, if the Wiener filter can be improved, the performance of BM3D can also be improved, since it uses the Wiener filter as one of its fundamental components.
In this study, we will primarily focus on the improvement of Wiener filter. Then, we will use this improved Wiener filter in BM3D to improve its response as well. Our objective is to eventually improve the denoising performance of all profiles of BM3D through improving the Wiener filter. Note that the authors previously published their preliminary idea of how the Wiener filter can be improved [9]. In this article, the authors will utilize their previous Wiener improvement idea to further improve the BM3D filtering scheme. In addition, we will also design some additional components to improve the performance of BM3D, especially the performance of BM3D profile. It is worth mentioning that from now on till the end of the article, we will refer to Additive White Gaussian Noise (AWGN) whenever the term noise is used.
The rest of the article is organized as follows. In Section 2, we will discuss the working procedure and parameterized setup of BM3D in details. In Section 3, we will discuss the Wiener filter and its variants. In Section 4, we will address the existing problems of the Wiener filter and BM3D that we are interested to solve in this study. We will propose our methodologies in Section 5 and report their performance in Section 6. Finally, we will conclude in Section 7 discussing some possible future work.
2 Block matching and 3D (BM3D) filtering
In recent years, probably the most discussed denoising technique is block matching and 3D (BM3D) filtering [5]. It was first suggested by Dabov et al. in 2007. Later, it was rigorously reviewed by Lebrun [12]. The idea has become extensively popular in denoising over the last few years. BM3D achieves excellent performance for reducing AWGN noise. In this section, we will discuss BM3D and its different profiles.
2.1 Algorithm of BM3D
The BM3D algorithm can be simply described in a step by step fashion. Let us start with Fig. 1 that shows the block diagram of BM3D [5].
BM3D takes the concept of nonlocal means (NLM) published in 2005 [2] in the sense that it also attempts denoising based on finding similar patches within a given window. BM3D has two identical steps namely step 1 and step 2. They are identical in the sense that they have no operational difference, rather the difference lies in the component that are used during the two steps. For example, the first step uses hard thresholding while the second step uses Wiener filtering. Other than that, both steps are identical. BM3D basically tries to denoise the noisy image in the first step to generate a basic estimate. This basic estimate is used in Wiener filtering of the second step as an oracle (i.e., degradation model) [5].
2.1.1 BM3D first step
In step 1, BM3D starts its operation by dividing the noisy image into a number of blocks or patches. For each patch, it then generates a window centering the block being processed. BM3D defines this center patch as a reference patch. Then, within this window, BM3D looks for the patches similar to the reference patch. Usually, a good number of similar patches are found. BM3D defines a threshold that decides whether two patches are similar or not (see Table 1). Once the similar patches are found, BM3D stacks all the similar patches together thus building a 3D block, where the first patch is the reference patch and others patches are sorted according to their distance to the reference patch. BM3D allows a number of 3D transform techniques to transform the domain from spatial to frequency (as indicated by the 3D transform in Fig. 1). After the 3D transform is performed, the most important part of the first step, known as hard thresholding, is executed. Hard thresholding is a filter that allows any coefficient with absolute value above a certain threshold to pass through, while converts the remaining coefficients to zero. This is the only operation in the first step that performs denoising, the rest are only to make a platform for hard thresholding. After hard thresholding, BM3D tries to generate a basic estimate. For this, it needs to transform the block coefficients to intensity values in the spatial domain. This is known as inverse 3D transform. After performing the inverse 3D transform, what is obtained is the 3D block that we started working with. But this time, it is denoised. Now, for each patch in the 3D block, an aggregation to estimate the reference patch is made. This aggregation is simply taking different weights and estimating each pixel. Once the aggregation is done, the basic estimate is ready to start the second step.
In theory, it is obvious that the more patches are present in the 3D block, the better estimates will be found for one single pixel as well as the better denoised basic estimates. However, according to Dabov et al. [5], in practice, it is seen that after a certain number of similar patches, BM3D does not seem to perform better.
2.1.2 BM3D second step
The second step is similar to the first step with two small differences. First, the 3D grouping is now performed on a basic estimate that was obtained from the first step, not on the noisy image as in step 1. Second, the hard thresholding is not used any more after the 3D transform. Instead, a Wiener filter is now used. We will discuss in Section 3 that the Wiener filter needs a degradation function H() to work. In BM3D, the 3D group built on the basic estimate is considered as the degradation function for BM3D while the corresponding 3D group of the noisy image is the degraded image function G().
Equation 1 shows how the Wiener filter works in BM3D. Here, \(\mathbb {P}^{basic}(P)(\xi)\) is the 3D block from the basic image and \(\mathbb {P}(P)\) is the corresponding 3D block from the noisy image. \(\tau ^{wien}_{3D}\) denotes the 3D transformation for the Wiener filter phase. Once the inverse 3D transform of Eq. 1 is computed, \(\mathbb {P}^{wien}(P)\) is found which is the final estimate for one block. Once the estimate is obtained, a weighted aggregation operation, like in step 1, is performed to build the final denoised image.
2.1.3 Parameters and profiles of BM3D
The performance of BM3D varies depending on the type of transformation used in both steps. There are two major profiles of BM3D, namely: DConly profile and wavelet profile. In DCOnly profile, BM3D uses a threedimensional discrete cosine transform as a 3D transform. For the wavelet profile, on the other hand, BM3D uses a combination of 2D biorthogonal transform and 1DHaar or WalshHadamard transform. DConly profile generally produces poorer results as compared to its wavelet counterpart [5, 6]. Wavelet profile may be defined as the mainstream BM3D since the authors of BM3D [5] recommended to use Wavelet transform in their proposed denoising method. This is because, the main stream BM3D has PSNR gain much better than DConly profile. For the wavelet profile, we used an exactly same parameterized setup as in BM3D [5]. We present the basic parameterized setup from the original article [5] in Table 1 for readersâ€™ convenience. We use exactly the same parameters to ensure the same environment for the experiment. The wavelet profile uses two subprofiles called normal profile and fast profile. The only difference between them is that the denoising performance is compromised in order to reduce the computational complexity in fast profile. Another difference between these two profiles is the fast profile uses predictive searching in order to decrease its searching time while the normal profile uses only exhaustive searching.
3 Wiener filter revisited
A Wiener filter provides an opportunity to deal with both noises and degradation. This feature makes the Wiener filter unique in both image denoising and restoration. This filter is also called minimum mean square error. This is because the core idea behind the Wiener filter is to satisfy an objective function which is the mean square error (MSE). In other words, this guarantees that the image restored by the Wiener filter \(\hat {f}\) will have minimum MSE with respect to original uncorrupted image f. Equation 2 shows that the expectation of a Wiener filter is to have minimum MSE between f and \(\hat {f}\).
The Wiener filter is defined by Eq. 3. Note that all terms are given in the transformed domain. Here, H(u,v) is the degradation function, H^{âˆ—}(u,v) is the conjugate complex of H(u,v), S_{ n } is the power spectrum of noise defined as \(S_{n}=\left {N(u,v)} \right ^{2}\), and S_{ f } is the power spectrum of undegraded image defined as \(S_{f}=\left {F(u,v)} \right ^{2}\). G(u,v) is the transform of the degraded image, and \(\hat {F}(u,v)\) is the final estimate for the restored image. Once the inverse transform of \(\hat {F}(u,v)\) is computed, \(\hat {f}(x,y)\) is obtained which is the denoised/restored approximation for original image f(x,y).
Comparing Eq. 1 with Eq. 3, it is evident that both equations are exactly the same, except that Eq. 1 works with 3D data. Equation 3 can be solved for G(u,v) as in Eq. 4 and rewrite Eq. 4 as in Eq. 5.
Now, if the noise is zero, the term inside the square brackets in Eq. 5 becomes 1, which means the Wiener filter is reduced to an inverse filter and works for only restoration. However, if there is noise, the Wiener filter incorporates itself for removal of noise along with restoration. This is what makes the Wiener filter unique.
4 Existing problems with Wiener filter and BM3D
4.1 Wiener filter objective function
4.1.1 MSEoptimized Wiener filter
As stated in Section 3, a Wiener filter tries to minimize its objective function shown in Eq. 2 while denoising/restoring a degraded image. This function is also known as mean square error (MSE) as defined in Eq. 6.
In this Equation, I and \(\hat I\) are considered as the true (or noise free) image and the reconstructed (or denoised) image, respectively. As the difference between these two gets smaller the closer the images are. Also, increased closeness indicates a more accurately denoised image. With this fundamental property, MSE is being used as an image quality metric [13].
The Wiener filter (Eq. 3) guarantees that the denoised/restored image is the closest image possible to the true undegraded image, since this filter is optimized for MSE. We might conclude at this point that in the bestcase scenario the Wiener filter may reconstruct an image whose MSE with respect to the true image is zero. That is, both images are exactly the same. However, there are still problems with the typical use of MSE that prevent the Wiener filter from achieving more accurate and perfect results.
4.1.2 Structural similarity
Wang et al. [20] showed that although MSE is a good quality measure, it is sometimes seriously misleading. This is because MSE does not consider anything other than the pointtopoint distance [21]. For example, if we add a constant value to all the pixel values of an image just to increase its brightness, the images are visually exactly the same; however, MSE still generates a huge error because of pointtopoint distance measurement. Figure 2 shows such an example of misleading of MSE measure. In this figure, we used the wellknown Lena image in Fig. 2a and added a constant 30 to all of its pixel values in order to increase its brightness. The brightened image is shown in Fig. 2b. Although there is no visual distortion in the image, the calculated MSE between them is 900. This value is not negligible, and hence MSE cannot be considered as a true error measure.
Now, let us also consider Fig. 2c. This is similar to Fig. 2b except that now we have not just added a positive constant to all brightness levels, instead, if the brightness value is less than 150, we subtracted 30 from it; otherwise, we added 30. Due to the presence of the square term in Eq. 6, the MSE still generates the same error value (900) as in Fig. 2, even though the image in Fig. 2c is greatly distorted, if it is compared with the image in Fig. 2b.
To avoid such misleading error situations, Wang et al. [20] proposed that modern image quality measurement metrics should not depend only on pointtopoint distances, it should also consider the geometric or structural similarity between the images. Otherwise, there is a possibility to have the same error results for different images as depicted above. Thus, they proposed a new error measurement metric called structural similarity (SSIM). Their proposed error measure, SSIM, is defined in Eq. 7.
In this Equation, x and y denote two blocks of the same position from the true image and the reconstructed image. Î¼_{ x } and Î¼_{ y } denote the arithmetic mean or average of x and y blocks, respectively. \(\sigma _{x}^{2}\) and \(\sigma _{y}^{2}\) indicate the variance of x and y blocks, respectively, while Ïƒ_{ x }y is the covariance between x and y. c_{1} and c_{2} are two constants to stabilize the division with weak denominator and their values are calculated using Eqs. 8 and 9, respectively. Here, L is the dynamic range of the image while k_{1} and k_{2} are two constants whose values are k_{1}=0.01 and k_{2}=0.03, respectively [20, 21].
This measure is calculated block by block. Therefore, a mean SSIM of all blocks is used to represent the final SSIM index value for the whole image. The SSIM index generally varies between âˆ’â€‰1 to +â€‰1 which is often taken as absolute to avoid a negativity value. Thus, the SSIM index is a fractional number between 0 and 1 (inclusive), where a 0 indicates no similarity and a 1 indicates exact similarity between two images. Using this measure, we have 0.9650 and 0.7850 for Fig. 2b, c, respectively, which indicates that Fig. 2c is much more distorted than Fig. 2b. This distortion was overlooked by MSE.
There are other misleading measures of MSE; however, we will not cover them in this study. For a detail list of misleading characteristics of MSE, we refer the reader to the literature [10, 20â€“22].
4.2 Existing problems of BM3D
4.2.1 Higher noise levels
BM3D has poor denoising performance to images corrupted by higher levels of noise as compared to its response to images corrupted by lower levels of noise. In other words, BM3Dâ€™s performance decreases as the noise level increases.
4.2.2 Poor performance of BM3D for DConly profile
BM3D requires the hard thresholding to be performed in a transform domain, after the 3D transform is performed on the similar image blocks. This 3D transform can be either a 2D+1D transform or a 3D transform. In BM3D, usually a 3D Wavelet is used as 3D transform or a combination of 2DDCT plus a 1D wavelet transform. However, the transform choice is not restricted. If the 3DDCT is used, this profile is called DConly profile. Note that, a profile is basically a set of parameters that is used in BM3D. In the DConly profile, coefficients are categorized into two categories: DC and AC coefficients, where the DC coefficient preserves the average of the block intensity, which is a significant piece of block information. It is worth mentioning that the DC coefficient might also possess some noise. When BM3D uses hard thresholding to get rid of the noise in the transform domain, it does not really treat the AC and the DC coefficients differently. As a result of this, the final outcome of DConly profile is poorer than other profiles (e.g., wavelet profiles).
5 Proposed method
5.1 SSIMoptimized Wiener filter
From the discussion in Section 4.1, it is evident that the MSE is not adequate for assessing the closeness between two images, it is rather good at assessing the distance between them. Instead, the SSIM is a more acceptable alternative. This is because, MSE deals with image data while SSIM deals with image information.
It should be noted that the idea of optimizing the Wiener filter with other quality measurement objective functions were tested by a number of objective functions such as sum of absolute difference (SAD) and median of absolute difference (MAD). However, since most of the similarity measures are rather closeness measures based on differences of pixelsâ€™ intensities and not on visual similarities, they could not actually further optimize the Wiener filter. Therefore, the choosing of SSIM as an objective function was logical.
The optimization of a denoising filter by SSIM instead of MSE is not very recent. Channappayya et al. in [3, 4] showed that any linear filter can be optimized by SSIM. They compared their proposed SSIMoptimized filterâ€™s result with the MSEoptimized Wiener filter. Their reported results showed that they were able to achieve a higher SSIM value compared to the MSEoptimized Wiener filter; however, their PSNR gain was still poorer than the MSEoptimized Wiener filter.
An SSIMoptimized Wiener filter, which we proposed earlier [9], considers the structural similarity between the reconstructed image and the true image. Since higher SSIM index indicates more similar images, our proposed Wiener filterâ€™s target is to estimate an image which has maximum SSIM possible. In this case, the expectation function of the Wiener filter becomes Eq. 10 which now needs to be maximized. In Eq. 10, E is the expectation and f and \(\hat {f}\) are the true and reconstructed images, respectively.
Having defined the expectation function, our target is to ensure that our designed Wiener filter maximizes our expectation. Taking a careful look at Eq. 3, we realize that replacing the term \(\frac {S_{n}}{S_{f}}\) by a variable K is reasonable and finding a suitable value of K is possible [7]. Therefore, for our proposed case, we can start with the lowest possible value of K and loop through the highest possible value of K. For each K, we record the SSIM error measure in a vector and then restore the image using that K for which the error has been recorded maximum. Thus, for a given range of noise level, it is guaranteed that our proposed Wiener filter should be SSIM optimized. Likewise, it should also provide better denoising and restoration.
Since the core of our proposed improvement over BM3D is the SSIMoptimized Wiener filter, interested readers may want to compare the performance given by both SSIM and MSEoptimized Wiener filter. We refer the reader to our previously published work [9].
5.2 3D zigzag thresholding
In the discrete cosine transform, the first coefficient is basically the average of all pixel values within a given block [17]. Therefore, for a precise inverse transforming result, the accurate DC coefficient is crucial. In order to make sure that the inverse transformation result is the product of vital block information, our proposed method does not apply hard thresholding on the DC coefficient. Instead, it only applies hard thresholding on AC coefficients. The AC coefficients carry various block frequency information [1, 7, 17]. This information varies from lowfrequency to highfrequency information. Thresholding all AC coefficients in the same manner might lead to losing some significant information, while preserving some other insignificant information.
In order to keep the most meaningful block information and just reduce the noise, we should use a zigzag thresholding, instead of a hard thresholding. Zigzag thresholding is realized by applying little or no thresholding on the DC coefficient and first few AC coefficients and then applying an increasingly higher thresholding on the rest of the AC coefficients (the higher the frequency, the more thresholding is applied). Determining the actual zigzag thresholding value can be chosen using a gamma curve. A gamma curve is as simple as Î»_{3D}=Îº^{Î³}, where Îº is the coefficient value, Î³ is a positive value â‰¥â€‰1 that is directly proportional to the coefficient number, and Î»_{3D} is the thresholded value. When Î³=1, the relation becomes linear. Using this thresholding scheme, we gradually increase the thresholding effect with the increase of the coefficient number. Note that this 3D zigzag thresholding proposal applies to only DConly profile of BM3D and not to the actual wavelet profile.
6 Results and discussion
6.1 Data set and parameterized setup
We used eight standard gray scale test images for our experiment. The images used are shown in Fig. 3.
For all these images, we recorded the responses of a MSEoptimized Wiener filter and our proposed SSIMoptimized Wiener filter. All performance comparison tables reported in this article are based on the average performance on these eight test images for each noise level. The noise level is generated using Eq. 11 where the value of Ïƒ is varied from 0 (minimum) to 100 (maximum) with Î¼=0 (zeromean). Here, Ïƒ is known as noise standard deviation. All other parameters are used with their default values as suggested by Dabov et al. [5] and presented in Table 1.
Experimental results presented in this article are reported in both subjective and objective forms. We used peak signal to noise ratio (PSNR) and structural similarity (SSIM) as our objective measure. As for the subjective assessment, the output from BM3D and from our proposed method is used to visually comapre the performance. While we present the result of all noise level experiments in the objective measure, we only present one noise level in the subjective assessment. We did so in order to keep the number of pages within the allowable limit.
6.2 Performance analysis of wavelet profile
6.2.1 Normal profile
Since the Wavelet profile itself exploits higher performance as compared to the DConly profile, even a small increase in PSNR/SSIM indicates a reasonable improvement. The experimental results for this profile are given in Table 2. The subjective measure is given in Fig. 4 for Lena image with Ïƒ=50.
6.2.2 Fast profile
The fast profile is similar to the normal profile, except that this profile is faster, as the identification of similar blocks is predicted. The performance of the fast profile is slightly lower than normal profile. The experimental results obtained for fast profile is presented in Table 3. Figure 5 shows a subjective measure for the Lena image with Ïƒ=50.
6.3 Extension of wavelet profile for color image denoising
One approach to employ Color BM3D (CBM3D) to 24bit true color images is to apply BM3D separately on each of its channels. However, correct grouping is one of the key properties of BM3D and it largely depends on the noise level. Again, grouping is a time consuming operation, doing it thrice makes the algorithm considerably slower. Moreover, three channels should generate three different groupings with sparsity of image information which lead to erroneous hard thresholding. BM3D extension to color image denoising is realized by converting a noisy RGB image into a luminance and chrominance transformed space. In this transformed space, the luminance signal contains most of the image information while the chrominance signals contain lowfrequency information. Therefore, CBM3D performs grouping on luminance channel only and uses exactly the same grouping for chrominance channels. The idea behind this form of grouping is that if the luminance of two blocks are mutually similar, then the chrominance of these blocks are also mutually similar [5].
We used the same concept for color image denoising as in CBM3D except that in step 2, we used our improved Wiener filter instead of the existing Wiener filter. The experiments are performed for both normal and fast profiles. The experimental results for normal color profile are presented in Table 4. For visual inspection of denoised color images of normal profile of Lena image with Ïƒ=50, we refer the reader to Fig. 6.
For the fast color profile, the experimental results are shown in Table 5. The output is shown in Fig. 7 for Lena image with Ïƒ=50 for a visual inspection of the reader. Since the fast profile compromises the performance in order to reduce time, the output is poorer than the normal profile. Therefore, for the image presented for Ïƒ=50 in Fig. 7, it may not depict a visible difference; however, this will be evident at higher noise levels.
6.4 Performance analysis of DConly profile
Figure 8 shows a comparison between the performance of the proposed zigzag thresholded result and that of BM3D for DConly profile, where Fig. 8a is the original image, Fig. 8b is the noisy image (Ïƒ=20), and Fig. 8c, d are the output produced by BM3D (DConly) and the proposed method, respectively. The results show an improvement in the quality of the denoised image when the proposed method is applied, especially at the face and shoulder areas. For further experimental results on the performance of DConly profile, readers are referred to Hasan [8].
6.5 Discussion
So far, we have presented a portion of our objective and subjective results from our experimentation. We have also compared our proposed method with the original BM3D. Objectively, our proposed method was able to produce higher PSNR and SSIM values than that of the original BM3D method. Tables 2, 3, 4, and 5 show that as the amount of noise increases, the proposed method managed to produce better denoising output, when compared to the original BM3D. For monochrome images, the SSIM gain by the proposed method was as high as 20% while the PSNR gain was as high as 4.49%. For color images, we achieved slightly less, as the maximum SSIM and PSNR gain were 6.63 and 2.56%, respectively. Subjectively, the quality of the denoising by the proposed method was better than that by the original BM3D. This was especially visible in the flat areas of the images, e.g., Lenaâ€™s shoulder and face. It is evident from the results that our proposed method achieved better objective and subjective denoising quality.
7 Conclusions
In this study, we rigorously reviewed the current stateoftheart image denoising scheme (BM3D) as well as its core component, the Wiener filter. We proposed an improved Wiener filter optimized for SSIM that has essentially improved the performance of BM3D. Through a number of experiments, we proved that our proposed Wiener filter, when used in BM3D, is capable of achieving highquality image denoising. Our idea works for both monochrome (gray scale) and color images. As a brief summary, our novel contributions in this study are the following: (1) reviewing the stateoftheart image denoising method BM3D with its components and profiles, (2) finding its existing shortcomings, (3) suggesting an improved Wiener filter optimized for SSIM [9], (4) using the SSIMoptimized Wiener in BM3D as its core component, and (5) thereby proving that the performance of original BM3D has been significantly improved in our proposed method (through detailed comparative studies via calculating a number of performance measurement metrics). (6) In addition, we have also proposed a technique named 3D zigzag thresholding for improving the poor performance of DConly profile of BM3D. All the innovations are discussed in detail in Section 5 while the performed tests, measurement metrics, and obtained results are discussed in Section 6.
7.1 Future direction
There are other variants of recently proposed SSIM called multiscale structural similarity (MSSSIM) [19]. Also, there are other image quality measurement metrics such as image quality index (IQI) [18], Normalized Correlation [11], Sum of Absolute Differences, and many others [15]. A possible future work will test our proposed method by using all of these measures. Recently, BM3D has been extended for video denoising. Also there is BM4D [14]. Since our idea is to change the core of BMxD in general, a study of assessing our method in all BMxD versions will be considered. In addition, there are recent works such as finding visually similar images using a convolutional deep neural network [24], it would be interesting to study if finding a similarity between images with deep neural networks helps in forming a quality metric that can eventually be used in a Wiener filter. Also, applying our improved denoising method could help in reducing noise for applications such as text extraction from complex background images [25].
Abbreviations
 AWGN:

Additive white Gaussian noise
 BM3D:

Block matching and 3D filtering
 BM4D:

Block matching and 4D filtering
 BMxD:

Block matching and xD filtering
 CBM3D:

Color block matching and 3D filtering
 DCT:

Discrete cosine transform
 IQI:

Image quality index
 MAD:

Median of absolute difference
 MSSSIM:

Multiscale structural similarity
 MSE:

Mean square error
 NLM:

Nonlocal means
 PSNR:

Peak signal to noise ratio
 RGB:

Redgreenblue
 SAD:

Sum of absolute difference
 SSIM:

Structural similarity
References
N Ahmed, T Natarajan, K Rao, Discrete cosine transform. IEEE Trans. Comput.C23(1), 90â€“93 (1974).
A Buades, B Coll, JM Morel, A nonlocal algorithm for image denoising. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPRâ€™05).2:, 60â€“65 (2005).
S Channappayya, A Bovik, R Heath, A linear estimator optimized for the structural similarity index and its application to image denoising. Int. Conf. Image Process. (ICIPâ€™06).2637â€“2640 (2006).
S Channappayya, A Bovik, C Caramanis, R Heath, Design of linear equalizers optimized for the structural similarity index. IEEE Trans. Image Process. 17(6), 857â€“872 (2008).
K Dabov, A Foi, V Katkovnik, K Egiazarian, Image denoising by sparse 3D transformdomain collaborative filtering. IEEE Trans. Image Process. 16(8), 2080â€“2095 (2007).
K Dabov, A Foi, V Katkovnik, K Egiazarian, BM3D image denoising with shapeadaptive principal component analysis. Signal Process. Adapt. Sparse Struct. Representations (SPARSâ€™09). (2009).
R Gonzalez, R Woods, Digital Image processing, 4th Ed (Pearson, New York, 2018).
M Hasan, BM3D image denoising using SSIM optimized Wiener filter (Masters Dissertation, Western University, 2014).
M Hasan, M ElSakka, Structural similarity optimized Wiener filter: a way to fight image noise. Int. Conf. Image Anal. Recognit. Lect. Notes Comput. Sci.9164:, 60â€“68 (2015).
A Hore, D Ziou, Image quality metrics: PSNR vs. SSIM. Int. Conf. Pattern Recognit. (ICPRâ€™10).2366â€“2369 (2010).
V Laparra, J MuÃ±ozMarÃ, J Malo, Divisive normalization image quality metric revisited. J. Opt. Soc. Am. A.27(4), 852â€“864 (2010).
M Lebrun, An analysis and implementation of the BM3D image denoising method. Image Process On Line. 2:, 175â€“213 (2012).
E Lehmann, G Casella, Theory of point estimation, 2nd edn (SpringerVerlag, New York, Berlin Heidelberg, 1998).
M Maggioni, V Katkovnik, K Egiazarian, A Foi, Nonlocal transformdomain filter for volumetric data denoising and reconstruction. IEEE Trans. Image Process. 22(1), 119â€“133 (2013).
A Mittal, A Moorthy, A Bovik, Noreference image quality assessment in the spatial domain. IEEE Trans. Image Process. 21(12), 4695â€“4708 (2012).
Vaseghi S, Advanced digital signal processing and noise reduction, 4th Ed (Wiley, Chichester, 2009).
G Wallace, The JPEG still picture compression standard. IEEE Trans. Consum. Electron. 38(1), xviiiâ€“xxxiv (1992).
Z Wang, A Bovik, A universal image quality index. IEEE Signal Process. Lett. 9(3), 81â€“84 (2002).
Z Wang, E Simoncelli, A Bovik, Multiscale structural similarity for image quality assessment. Asilomar Conf. Signals, Syst. Comput.2:, 1398â€“1402 (2003).
Z Wang, A Bovik, H Sheikh, E Simoncelli, Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600â€“612 (2004).
Z Wang, A Bovik, Modern image quality assessment: Synthesis Lectures on Image, Video, and Multimedia Processing (Morgan and Claypool Publishers, San Rafael, 2006).
Z Wang, A Bovik, Mean squared error: love it or leave it? A new look at signal fidelity measures. IEEE Signal Proc. Mag. 26(1), 98â€“117 (2009).
N Wiener, Extrapolation, interpolation, and smoothing of stationary time series (MIT press Cambridge, Cambridge, 1964).
C Yan, H Xie, D Yang, J Yin, Y Zhang, Q Dai, Supervised hash coding with deep neural network for environment perception of intelligent vehicles. IEEE Trans. Intell. Transp. Syst. 19(1), 284â€“95 (2018).
C Yan, H Xie, S Liu, J Yin, Y Zhang, Q Dai, Effective Uyghur language text detection in complex background images for traffic prompt identification. IEEE Trans. Intell. Transp. Syst. 19(1), 220â€“229 (2018).
Acknowledgements
Not applicable.
Funding
This research is partially funded by the Natural Sciences and Engineering Research Council of Canada (NSERC). This support is greatly appreciated.
Availability of data and materials
The datasets supporting the conclusions of this article are included within the article.
Author information
Authors and Affiliations
Contributions
MES suggested the idea. MH devised the algorithm based on that and carried out the experiments. MES continuously supervised the experiments and suggested modifications. MH prepared the manuscript draft, and MES edited and proof read the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Hasan, M., ElSakka, M. Improved BM3D image denoising using SSIMoptimized Wiener filter. J Image Video Proc. 2018, 25 (2018). https://doi.org/10.1186/s136400180264z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s136400180264z
Keywords
 Image denoising
 Image restoration
 BM3D
 Wiener filter
 Structural similarity
 Collaborative filtering
 Hard thresholding
 Mean square error