AN ENHANCEMENT OF EDGE PRESERVATION IN OAMNHA DENOISING USING TEXTURE BOUNDARIES

471-489


1.INTRODUCTION
Image denoising also known as noise reduction is the process of eradicating the noise from an image during edge preservation process. Normally, edge preserving process is one of the image processing techniques that remove noise or textures while preserving sharp edges.
In these days, digital images play the most significant role in different applications such as magnetic satellite box, computer tomography, resonance imaging, and ecological systems.
In general, the images are captured by the sensors. In some cases, like faulty components, data acquisition problems and normal phenomena interference, the most significant data can be degraded. As well, the noise may be occurred by the inaccurate compression and transmission. As a result, image denoising process is needed as primary process in the image processing to compensate the data corruption. Still, image denoising technique has lot of challenges since the artifacts are occurred during noise removal which blurs the images. Hosotani et al. (2015) introduced an image denoising process with edge preserving and segmentation based on the Mask NHA (M-NHA). In this process, the zero-mean white Gaussian noise was eliminated by using the high-resolution frequency analysis. The regions including identical texture were analyzed on the noisy image. The non-uniform regions obtained by segmentation process were analyzed by using M-NHA to improve the Peak Signal-To-Noise Ratio (PSNR). Conversely, an optimization of the parameters used in the segmentation was not efficient and the threshold value for edge detection was fixed to suppress the unwanted data. Hence, an OAMNHA based image denoising technique was proposed by using SVM and firefly optimization algorithms. In this technique, Support Vector Machine (SVM) was applied for learning the parameters used in the segmentation and firefly algorithm was used for optimizing the threshold values for many noisy images. Therefore, the boundary distortion is reduced by defining the texture boundaries due to segmentation and detection process. Also, the accuracy of the edge-preserving segmentation and detection is efficiently increased.
The remaining part of the article is organized as follows: Section II presents different image denoising approaches which are related to the proposed technique. Section III clarifies the methodology of proposed image denoising technique. Section IV illustrates the experimental results and Section V concludes the entire discussion. Chen et al. (2013) proposed an edge preserving image denoising with a closed form solution.
In this method, a novel pixel-based algorithm was proposed that formulates the image denoising problem as the Maximum-A-Posterior (MAP) estimation problem by using Markov Random Fields (MRF). This was converted into a continuous label assignment problem based on a Gaussian MRF model and then a closed form globally optimal solution was obtained. The pre-estimated image edge information was added into the energy function construction to preserve the image structures. Moreover, patch similarity based pairwise interaction was performed to better preserve the image details and achieve high robustness. However, Mean Squared Error (MSE) of this algorithm was high. Jiao and Huang (2014) proposed a new wavelet packet transform adaptive threshold image denoising method based on edge detection. By edge detection method, the wavelet packet coefficients corresponding to detected edge and other non-edge wavelet packet coefficients were treated by different threshold. By using the relativity among wavelet packet coefficients and neighbor dependency relation, the new variance neighbor estimate method was adopted and then the adaptive threshold was generated. On the other hand, the computational complexity of this method was high. Gao, Wang and Liu (2015) proposed an image denoising method based on edge detection and pre-thresholding wiener filtering of multi-wavelets fusion for removing Gaussian noise from digital images. Initially, the noisy image was decomposed by using multiple wavelets and the edge of image was detected through wavelet multi-scale edge detection. Based on this, the wavelet coefficients belonging to the edge position were dealt with the improved wavelet threshold method whereas the others were dealt with the pre-thresholding wiener filtering. At last, the fusion algorithm was used based on wavelet analysis for obtaining the denoised image. However, the computational burden i.e., computation time was high and it requires an adaptive threshold selection using advanced optimization methods to further improve the image denoising performance. Guo, Zhang and Zhang (2018) proposed an edge-preserved image denoising algorithm based on local adaptive regularization which can adaptively adjust denoising in accordance with different regions of noisy image by containing residual local energy function. In this method, detailed information of image was also well preserved at the process of denoising. Jain and Tyagi (2017) proposed an adaptive edge-preserving image denoising technique by using patch-based weighted-Singular Value Decomposition (SVD) filtering in wavelet domain. In this filtering method, a group of non-local similar patches were considered for each local patch in the subband and the weights were estimated for this patch group based on the significance of its singular values. After that, these weights were used as a threshold for SVD filtering of the considered patch group. This process was applied to each local patch and all patches were aggregated together for obtaining the thresholder results in the subband. However, this algorithm was not able to obtain the depth of the image which has the shadow effect and the intensity values ranged in the black to gray.
Zhao and Shang (2018) proposed an adaptive edge-detection method i.e., a method of iterative threshold edge detection based on histogram. In this method, multi-scale wavelet transform was used for preprocessing the image in which the image detains were highlighted and the effect of manual setting filter coefficients was avoided. The variation of gray values between the pixels of local region was used for computing the gradients and the gradient directions were extended to four directions. The adaptive method was used for computing the threshold of edge-detection and the image was represented by histogram. After that, the ratio of the number of pixels in the bar and the total numbers of pixels were used for selecting the initial threshold.
The regions on both sides of the initial threshold were used for calculating the high threshold and low threshold iteratively. The detection errors, connection errors and the pseudo-edges caused by selecting the threshold artificially were avoided. However, the

MATERIALS AND METHODS
In this section, the proposed NF-OAMNHA-CT based image denoising technique using NF approach is explained briefly. Initially, the image i(x,y) with size M×N is transformed into frequency domain such as CT. Then, the image is given to the NF edge detector to detect the edge regions and homogenous textures as follows: The structure of the proposed Neuro Fuzzy edge detector shown in Figure 1(a) has four NF networks which are functioning as sub-detectors in the four directions, namely vertical, horizontal, right diagonal and left diagonal, respectively. Each sub-detector can operate on a window size of 3x3 which is shown in Figure 1(b). Also, each sub-detector estimates a different neighbourhood correlation between the center pixel of the filtering window and two of its neighbour's. Each NF sub-detector is a first-order Sugeno type Fuzzy Inference System (FIS) with 3-inputs and 1-output. Each input has 3 generalized bell type membership functions and the output has a linear membership function (m). The input-output correlation of any of the NF sub-detectors is as follows: Consider i_1,i_2,i_3 are the inputs of the NF sub-detector and e is its output. Each possible combination of inputs and their related membership functions is denoted by a rule in the rule base of the NF sub-detector. As the NF sub-detector consists of 3 inputs and each input has 3 membership functions, the rule base has 27 rules totally which are given below: if (i 1 is m 11 ) and (i 2 is m 21 ) and (i 3 is m 31 ), then R 1 = F 1 (i 1 , i 2 , i 3 ) if (i 1 is m 11 ) and (i 2 is m 21 ) and (i 3 is m 32 ), then R 2 = F 2 (i 1 , i 2 , i 3 ) if (i 1 is m 11 ) and (i 2 is m 21 ) and (i 3 is m 33 ), then R 3 = F 3 (i 1 , i 2 , i 3 ) if (i 1 is m 11 ) and (i 2 is m 22 ) and (i 3 is m 31 ), then R 4 = F 4 (i 1 , i 2 , i 3 ) · · · if (i 1 is m 13 ) and (i 2 is m 23 ) and (i 3 is m 33 ), then R 27 = F 27 (i 1 , i 2 , i 3 ) Here, m ij is the j th membership function of the i th input, R k is the output of the k th rule and F k is the k th output membership function. The input membership functions (m ij ) are generalized bell type which is defined as follows: (1) The output membership functions (Fk) are as: For i = j = 1,2,3 and k=1,2,…,27 The factors a, b, c and d in the above equations are constants that differentiate the shape of the membership functions. The optimal values of these parameters are determined by learning process. Each NF sub-detector is learned individually. The setup used for learning is shown in Figure 2.
Here, the parameters of the Neuro Fuzzy sub-detectors under learning are iteratively fine-tuned so that its output converges to the output of the ideal edge detector which can 3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254 -4143 properly detect the positions of the edge pixels in the given image. The ideal edge detector output represented by the target training image is used only for learning process. The threshold value is half of the available dynamic range for the pixel luminance values.
For 8-bit images where the pixel values range between 0 and 255, the threshold value is taken as 128. The input and output correlation of the postprocessor is described as follows: Consider e 1 , e 2 , … ,e 1 is the outputs of the NF sub-detectors where l is the number of NF sub-detectors used. The output of the postprocessor is computed in two steps: Step 1: The average value of the individual NF sub-detectors outputs is computed as: (5) Step 2: The computed e AV value is altered into 0 (black) or 255 (white) by relating it with the threshold as: In equation (6), e(x,y) represents the result of the postprocessor i.e., the output of the Neuro Fuzzy edge detector that denotes the conclusion whether the center pixel of the filtering window is an edge pixel or not. This is continued until all pixels of the noisy image are classified properly. Thus, the edges from the noisy image are extracted efficiently and then the edge regions are defined accurately. After that, the texture boundaries formed and OAMNHA technique is applied for each homogeneous texture region in order to reconstruct the noise-free images with the increased segmentation accuracy and reduced computational complexity.

PSNR
It is the ratio of the maximum signal power to noise power and computed as: The comparison of denoising PSNR which is made between proposed NF-OAMNHA-CT and existing OAMNHA-CT techniques is shown in

MAE
It is the measure used to calculating the average magnitude of error in the prediction groups and computed as: In equation (9), P i is the prediction value, Y i is the true value and ε i is the absolute error.

SSIM
It is the similarity measure between any two images i(x,y) and computed as follows: In equation (10), the averages of x and y are represented as μ x and μ y . Also, the variances of x and y are represented as σ x 2 and σ y 2 x,y respectively, c 1 and c 2 are constants. As well, the covariance of x and y is denoted as σ xy .

CONCLUSIONS
In this paper, NF-based edge-preserving segmentation is proposed to enhance the OAMNHA-based image denoising technique in different frequency transform domains.
Initially, the considered image is fed into the Neuro Fuzzy edge detector to classify the pixel of the given image is whether edge pixel or not. For all pixels in the image, Neuro Fuzzy filtering process is repetitive and the edges are extracted precisely from the noisy image.
Then, the extracted edges are defined as the edge regions. Also, the texture boundaries are constructed based on the homogeneous texture segmentation. For each segment, OAMNHA technique is applied to restore the noiseless image efficiently. Therefore, the time consumption for edge detection is reduced and the accuracy of the segmentation process is also increased. Finally, the experimental results proved that the proposed NF-OAMNHA-CT technique achieves higher PSNR, SSIM and less MAE than the OAMNHA-CT based image denoising technique. In many applications, e.g., medical or satellite imaging, the edges are key features and thus must be preserved sharp and undistorted in smoothing/ denoising. Also, edge preserving denoising is useful for the images capturing from camera which is an optical device and prone to sensor noise, especially in dark environments or environments with extreme high dynamic range.