Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: This paper proposes a machine vision scheme for denoising, feature space gradient preserving, and
Available online 3 June 2017 detecting weld defects in noisy weld X-radiography images; particularly, for the images that are in low-
contrast and contain noises. The detection of small weld defects present on noisy image is extremely
Keywords:
difficult in non-destructive testing through machine vision. The presence of high gradient magnitude and
Anisotropic diffusion
the low intensity in the feature space of a noisy image are the main characteristics of weld defects.
Image denoising
Weld defects These characteristics can be considered to refine and obtain noise-free images for detection of weld
Segmentation defects. This study proposes a modified anisotropic diffusion model, which considers a local probability
X-radiography image value of gray-level and an adaptive threshold parameter in diffusion coefficient function to adjust the
implication of low edge gradient of the feature space from the noisy image. Furthermore, an entropy
based stopping criterion has been introduced to terminate the diffusion process. This proposed model is
compared with the existing models, and its performance is evaluated through Mean Square Error (MSE),
Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), Entropy (E) and Mean Structural Similarity
(MSSIM) measures. Experimental results confirm the reliability of the proposed model.
© 2017 Elsevier Inc. All rights reserved.
1. Introduction that of the normal distribution, i.e., the values that the noise can
take on are Gaussian-distributed. In digital images, including those
Ultrasonic, Eddy current, and X-radiography inspection tech- in X-radiography, it is common practice amongst the researchers to
niques are commonly used in non-destructive evaluation industry. model the present noise as Gaussian and apply suitable denoising
Digital X-radiography has widely been used in Non-Destructive filters on the image for its quality improvement. Gaussian noise is
Testing (NDT) and particularly in weld defect detection [1]. The also called electronic noise because it arises during image digiti-
detection of defects depends on the quality of the digitized image, zation process, which is independent of original pixel values and
which is subjected to certain factors, such as noise, low contrast, interrupts the gray values during digitization. Moreover, the noisy
defects of different dimensions, and indiscernible defects in the digital weld X-radiography image reduces the ability to identify
image background. In experimental sciences, noise refers to any tiny weld defects, and it is further worsened if the image con-
random fluctuations of data that hinders perception of an expected taminated by noises. However, the features present in images are
signal. The noise is a summation of unwanted or disturbing energy degraded or contaminated by the unknown Gaussian noise levels.
from natural or man-made sources, and is typically distinguished Therefore, it is a challenging task to remove such noises with-
from interference or distortion. Sources such as automobiles, air- out disturbing the image features during noise removal process.
craft, ignition electric motors and switching gear, high voltage To devise a suitable model to remove the unknown Gaussian noise
wires and fluorescent lamps cause industrial noise, produced by present in the digital weld X-radiography images during denoising
the discharge present in all these operations. Gaussian noise is sta- and smoothing process therefore turns out to be important.
tistical noise having a Probability Density Function (PDF) equal to Incorrect feature extraction leads to detection of inaccurate de-
fects. Moreover, preserving the features present in the Region of
Interest (ROI) is considered as severe problem and it is a pre-
* Corresponding author. requisite for image interpretation and analysis during denoising
E-mail address: sgk@mca.sastra.edu (G. Sethumadhavan). and smoothing processes. In this regard, several denoising and
http://dx.doi.org/10.1016/j.dsp.2017.05.014
1051-2004/© 2017 Elsevier Inc. All rights reserved.
M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126 113
smoothing techniques of edge-preserving have been approached to Tian et al. [13] proposed another modified version of P–M
remove noises and revamp the features of an image without dis- model, a Difference Eigenvalue Perona Malik (DEPM) model, which
torting its region and region edges in the image. preserves fine features during noise removal processes. This model
Among the denoising and smoothing methods, nonlinear meth- was adapted and an edge indicator called difference eigenvalue
ods have been found to deliver more adequate results in com- to measure image gradient magnitude in the diffusion process
parison to linear methods. Several nonlinear approaches using was introduced. Surya Prasath and Vorotnikov [14] proposed a
Anisotropic Diffusion Filter (ADF), fourth order Partial Differential scheme for image denoising and restoration called weighted and
Equation (PDE), Non-Local Means (NLM), combination of domain well-balanced anisotropic diffusion. In order to remove noise with-
and range filters, nonlinear Gaussian filters and bilateral filters and out loss of fine details, a new variant of the flow is considered
their extensions to the trilateral filters [2,3], and are widely known. using weight parameter in the divergence diffusion process.
In this paper, an anisotropic diffusion based scheme is proposed to Another category includes edge-preserving approaches that are
solve the problem of weld defect detection in the noisy images. based on changing the diffusion function or introducing new oper-
Perona and Malik were the first to apply the anisotropic diffusion ations [15–23]. Barcelos et al. [15] approached the Well-Balanced
equation for image smoothing and edge detection. This classical Flow (WBF) based on Total Variation (TV), which is used in the
method is known as P–M Model [4], which has developed for divergence process of the P–M model. The Gaussian kernel was uti-
edge detection, enhancement of image, smoothing the image, im- lized as pre-smoothing to avoid noise oscillations. Mean curvature
age restoration, segmentation of features in the image. In [5,6], two flow equation and Nordstrom’s biased PDE were employed in WBF
approaches have been proposed followed by P–M model, which model. Coupled anisotropic diffusion [16] scheme was introduced
have motivated to introduce many approaches till date. Catte et al. using the PDE systems consisting of AD equations. This scheme
[5] proposed a scheme that carries out pre-denoising with a Gaus- will smooth regions during the diffusion function, where the gra-
sian filter before each iteration of diffusion function. This scheme dient is trivial, but will preserve it where the gradient is large on
can efficiently reduce noise; however, the selection of a suitable a noisy image. A kernel anisotropic diffusion (KAD) [17] method
value for σ parameter of Gaussian filter is still a problem. Rudin et utilizes a kernalized gradient operator in the diffusion function for
al. [6] proposed nonlinear total variation (TV) method. This scheme noise removal and detection of edge in a noisy image. To terminate
provides a better effect on staircases using gradients alone. the diffusion procedure automatically, the KAD introduces adap-
In recent years, several modified anisotropic diffusion (AD) tive diffusion threshold estimation using Median Absolute Devia-
models with different schemes have been approached for digital tion (MAD) of the gradient magnitude and employs Mean Absolute
image processing. The modified AD models can be classified into Error (MAE). Dai Fang et al. [18] proposed a nonlinear diffusion
two categories based on denoising process and edge preserving. model for smoothing and sharpening of image based on the Mean
Regarding the denoising methods that are using local information Curvature Motion (MCM) and the inverse heat diffusion equations.
for diffusion process, Chao and Tsai et al. proposed several ap- The staircase effect in the image has been considered in the Ramp
proaches [7–12] based on anisotropic diffusion for noise removal Preserving Perona–Malik model (RPPM) [19]. The RPPM model in-
and defect detection in sputtered glass, astronomical images, Thin troduces isotropic and anisotropic diffusion properties to preserve
Film Transistor images, solar wafer, and surface images. Tsai and staircase effect and edge region respectively. Guo et al. [20] pro-
Chao [7] formulated an anisotropic diffusion (TCAD) method for posed an Adaptive Peronal–Malik (APM) model for image denoising
defect detection in sputtered glass substrates with inhomogeneous based on variable exponent. The APM roughly segments the image
textures. This method employs a non-negative reducing function into two regions using the edge indicator. Among these regions,
with a strengthening gradient threshold as the diffusion coeffi- APM remove the noise and preserve the edge accordingly. Wang
cient to adjust adaptively the importance of edge gradients. Chao et al. [21] proposed a Modified Perona–Malik (MPM) model based
and Tsai [8] further proposed an improved anisotropic diffusion on directional Laplacian. In order to remove noise at the same time
method, known as CTAD. In order to eliminate the noisy stars from preserving edges, a weighted Laplacian was incorporated where di-
a nebula image, this method incorporated the local variance infor- rectional Laplacian diffuses image along the edge direction of the
mation of gray-levels in the diffusion coefficient function. Another original image. Maiseli and Gao [22] proposed a diffusion-driven
version of diffusion model [9] called CTAD-TFT has successfully iterative method to extract edges from noisy images. This method
sharpened and smoothened low-contrast surface images, particu- utilized anisotropic diffusion to suppress noise and introduced an
larly, in glass substrates used in Thin Film Transistor-Liquid Crystal edge-sensitive detector to highlight the edges of the feature. Xu et
Displays. Weighted sharpening diffusion function was employed in al. [23] have proposed a method based on semi-adaptive threshold
the classical P–M model for smoothing process. in anisotropic diffusion process. Gaussian filter used to distinguish
In [10], micro-cracks in polycrystalline solar wafer surface im- corrupted pixels and noise-free pixels as a pre-denoising process
ages involving heterogeneous texture patterns were detected using then in the post-processing, a semi-adaptive threshold function in
anisotropic diffusion. The normalized gray-level of an original im- diffusion coefficient is used.
age was used as diffusion coefficient function for smoothing pro- On weld X-radiography image processing domain, in particu-
cess. In this method, two post processing steps, binary threshold- lar denoising and smoothing process, very fewer researchers have
ing and morphological operations, were executed to segment the applied AD based approach in detecting weld defects [24–26].
defects after the smoothing of image. Generalized diffusion coeffi- Mhamed et al. [24] proposed a modified AD model for automatic
cient function [11] in anisotropic diffusion was proposed to detect defect detection in X-radiography images. This model initially con-
defects in low-contrast surface images that can adaptably adjust volutes the gradient image using median filter. Then the edge
the curve of the diffusion coefficient function. This method ap- sharpening function was added to the diffusion coefficient function
plies a stochastic evolutionary computation algorithm and particle for weld defect detection. Halim et al. [25] proposed a model based
swarm optimization (PSO) to determine the parameter values au- on PDE for weld defect detection on digital radiography image. This
tomatically. A modified anisotropic diffusion technique named CTD method utilized P–M model as a denoising process for smooth-
for an edge-preserving smoothing was proposed [12]. This method ing, and Chan–Vese method for segmentation and defect detection
preserves edges besides fine details during noise removing in the process. Dang et al. [26] considered the P–M anisotropic diffusion
diffusion process. It employs both local gradient and normalized filter as an image smoothing besides preserving the edges of the
variance of gray-level in diffusion coefficient function with adap- defects. In this approach, linear weighting and fuzzy enhancement
tive selection of threshold parameter. algorithms are involved in segmentation of weld defect.
114 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
(i ) 1
g (∇ I t (x, y ) = (i )
(6)
|∇ I (x, y )|
1 + [ k(0t)t −(1/3) ]2
1 1
2 anisotropic diffusion. In this paper, we propose a revised AD model
σt2 (x, y ) = I t (x + i , y + j ) − I t (x, y ) (9) that employs both probability value of local gray-level and gradi-
9
i =−1 j =−1 ent for denoising and smoothing the noisy image adaptively. The
proposed method not only preserves edges, but also retains the
where I t (x, y ) is the mean of the gray levels in the 3 × 3 neigh-
fine details of the feature during smoothing process. The proposed
borhood window at iteration t.
model is explained in the following Sections 3.1 and 3.2.
2.2.3. CTAD-TFT
The CTAD-TFT model [9] combines both the smoothing and 3.1. Reduction of noise
sharpening processes in a single function and the same has been
exhibited on surface images of glass substrates used in TFT-LCDs. A pixel at coordinate (x, y ) and at iteration t, the gray-level
In order to improve the subtle defects in the image, this model in- probability is computed from its 3 × 3 neighborhood region, S xy
corporates the weighted sharpening process in the diffusion func- as:
tion. The sharpening diffusion coefficient function ν (∇ I ) is defined S xy I t (x, y )
as pt (x, y ) = , (14)
I t (x, y )
ν (∇ I ) = α · 1 − g (∇ I ) (10) where I t (x, y ) denotes the sum of gray-levels in the 3 × 3 neigh-
where α is the weight of sharpening coefficient function with 0 ≤ borhood region S xy .
α ≤ 1. The diffusion coefficient function g (∇ I ) is the same as that Eq. (4) can be rewritten by integrating the Eq. (14) in the diffu-
defined in Eq. (4). sion function as:
The modified diffusion model of CTAD-TFT has been reformu-
(i ) S xy 1
lated as g ∇ I t (x, y ), pt (x, y ) = (i ) S xy
, (15)
∇ I t (x, y )· pt (x, y ) 2
1+( )
1
4
(i ) (i ) k
I t +1 (x, y ) = I t (x, y ) + g (∇ I t (x, y ) − ν (∇ I t (x, y ) where k is a positive fixed threshold used as a regulator to denoise
4
i =1 and also to improve the edge strength. Eq. (2) can be modified
× ∇ I t(i ) (x, y ) (11) as Eq. (16) using the proposed function of diffusion coefficient as
given in Eq. (15).
2.2.4. CTD model
The CTD model [12] was proposed to preserve edge regions and I t +1 (x, y )
1
(i )
detail information during noise reduction process. The CTD model 4
S xy (i )
includes both local gradient and normalized variance of gray-level. = I t (x, y ) + g ∇ I t (x, y ), pt (x, y ) · ∇ I t (x, y ) (16)
The diffusion coefficient equation is given as 4
i =1
(i ) 1 The new flux function of the proposed diffusion coefficient
g ∇ I t (x, y ), σt2, N (x, y ) = (12)
(i )
∇ I (x, y )·σt2, N (x, y ) 2 function given in Eq. (15) is formulated as follows:
1+[ t ]
k
S
The normalized variance is defined as follows: φ ∇ I t(i ) (x, y ), pt xy (x, y )
(i ) S (i )
σt2 (x, y ) − Min σt2 = g ∇ I t (x, y ), pt xy (x, y ) · ∇ I t (x, y ) (17)
σt2,N (x, y ) = 1 + · 254 (13)
Max σt2 − Min σt2 Fig. 3 illustrates the 3D plot of the diffusion coefficient function
(i ) S xy
2 2
where Min σ and Max σ are the minimum and maximum gray- of Eq. (15) with two variable functions ∇ I t (x, y ) and pt (x, y ).
t t
level variances of the diffused image at iteration t, respectively. It depicts that the values of the diffusion coefficient function
(i ) S xy
σt2 (x, y ) is the local gray-level variance (see Eq. (9)). g (∇ I t (x, y ), pt (x, y )) are increased suddenly when both the
116 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
Fig. 5. The comparison of various diffusion functions at sample 3 × 3 window: (a) added Gaussian noise image; (b) P–M model (iteration = 20 and k = 5); (c) CTAD model
with local variance (iteration = 50 and k = 0.05); (d) the proposed model with local probability (iteration = 12 and k = 5).
M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126 117
S xy
p 1 (x, y ) is not considered as it is computed before the start of the
diffusion process. Thus, the lower probability value produces the
S xy
higher diffusion coefficient as shown in Fig. 6. Therefore, pt (x, y )
in the proposed diffusion coefficient can sufficiently smoothen the
noisy image.
In weld X-radiography images, the weld defects are of various
dimensions with low gradient in large area of noisy background
and thus need processing. The low gradient present in the fea-
ture space of a noisy image will affect the accuracy of detecting
the features viz. defects. Here, we ensure that the feature space
that is weld defect is retained without any disruption during the
S xy
smoothing process. pt (x, y ) increases the gradient value and de-
noises the feature space. Moreover, the random selection of k value
S xy
Fig. 7. Graph of pt (x, y ) in various iterations t for the image in Fig. 5(a). cannot always be effective for the noise removal, since the resul-
tant image may become over or under smoothened. Therefore, in
this proposed model, an automatic selection of k has been intro-
duced. The following section illustrates an adaptive selection of k
and besides, a criterion proposed to terminate the diffusion func-
tion iterations.
Fig. 9. Smoothened images of Fig. 5(a) and 2D intensity profiles: (a) P–M model; (b) proposed model; (c) scan line 35 in Fig. 9(a); (d) scan line 35 in Fig. 9(b).
118 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
L −1
(t )
(t ) Table 1
St = − H i ln H i (18) The t and k parameter values.
i =0 Model No. of iteration (t ) Threshold parameter (k)
P–M 20 5
3.2.1. Adaptive selection of k
TCAD 150 5
Digitized images will have different entropy value based on its CTAD 50 0.05
existing noise levels. In this proposed model, the lower integral of CTAD-TFT 30 Adaptive selection of k using
Shannon entropy, S t =0 of the input image is initialized to k, a mean gradient magnitude
threshold parameter in the diffusion coefficient function. The as- CTD 150 60
3.2.2. Diffusion function iteration stopping criterion It has been found that a higher SNR value has been achieved in
In order to terminate the iteration process to obtain the de- the proposed model on the scan line with less iteration. Therefore,
sired smoothing during diffusion process, a stopping criterion is the adaptive selection of k and diffusion stopping criterion of the
proposed. The Shannon entropy value S t at iteration t is calculated proposed model have performed well yielding smoothened image.
on diffused image. The diffusion function iteration is terminated if
| S t −1 − S t | < 0.0005. 4. Image database, comparative experimentations and
Fig. 8 shows that the Shannon entropy value of each iteration discussions
of Fig. 5(a) by the proposed model. It can be seen that the en-
tropy value of the processed image decreases in each iteration
and stabilizes after a finite number of iterations based on the 4.1. Weld X-radiography image database
stopping criterion. In this experiment, diffusion process stops at
11th iteration where Shannon entropy values of the 10th iteration The X-radiography digital images of welds used in this paper
and 11th iteration are equal to three decimal places of accuracy were taken from GRIMA X-ray database [27] collected by Domingo
(S 10 = S 11 = 4.008).
Mery et al. in the interest of X-ray testing community. The test
images used for the analysis in this paper are taken from GRIMA
3.2.3. Performance analysis of adaptive selection of k and iteration
X-ray database and their acquisition standards are mentioned in
stopping criterion
their website as follows: The X-radiography acquired according to
In order to ascertain the appropriateness of adaptive selection
ISO 17636-1 and the optimized voltage is 150 kV for thickness of
of k and the stopping criterion of the proposed model, a compari-
8 mm to 20 mm steel. The weld X-radiography films have been
son analysis between the classical P–M and the proposed model is
digitized with the Lumisys LS85 SDR scanner with 12 bit pixel
experimented. The proposed model is applied on the image given
in Fig. 5(a) with adaptively selected threshold parameter k = 4 and depth. The original 12 bit data was rescaled to 8 bits with a linear
the process terminate at the 11th iteration and achieve the resul- LUT proportional to optical film density. The pixel size of the digi-
tant Signal-to-Noise Ratio (SNR) value 6.297. The same threshold tized X-ray radiography image is 40.3 μm (630 dpi). The image file
value has been used in the classical P–M model with 20 iterations format is TIFF without compression. The test images are 8 bit gray
and obtained the SNR value 6.181. The corresponding smoothened values.
images are given in Fig. 9(b) and Fig. 9(a). The SNR is computed In order to focus our attention on the region where the weld
using mean and standard deviation (S N R = μ/σ , where μ is the defects are available, we isolated a rectangular ROI manually that
mean and σ is the standard deviation of the considered pixel val- contains the defect under study from the full size image. Fig. 10
ues). Higher the value of SNR, the better will be the noise level. shows some of the test images with defects such as porosity, gas
A scan line that passes through the defective region is consid- pore, undercut, slag inclusion, tungsten inclusion, and crack which
ered from the smoothened images of both the models. Their 2D we have considered for experimentation. For performing the de-
intensity profiles of the scan lines are considered for visualization noising experiment, Gaussian noises of different standard devia-
and their SNR values are extracted to quantify and compare with tions (chosen randomly between 10% and 40% of σ ) were added
these models. These results are presented in Figs. 9(d) and 9(c). to these images.
M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126 119
Fig. 11. Smoothened results; Column (a), (b) and (c) first row: Noisy weld X-radiography images of Fig. 10(a), 10(b), and 10(c), respectively; from row II to VII: corresponding
smoothened images obtained from P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed model, respectively.
4.2. Implementation tion method) taken to test all the images were less than or equal
to 5. Therefore, the proposed model selects the k value appro-
The classical P–M [4], TCAD [7], CTAD [8], CTAD-TFT [9], and priately as ascertained by Section 2.1. Moreover the number of
CTD [12] models are implemented for smoothing the images, and iterations required to get the desired smoothing is less than or
the outcomes of these models were compared with the proposed close to minimum number of iterations performed in the existing
model. Table 1 shows the various values of the parameters that models besides the number of iterations required to stop the pro-
were taken in the existing models during implementation. Figs. 11 cess is also dynamically adaptive. We present the analysis of the
and 12 illustrates the various smoothened results obtained from segmentation and quality performance of the smoothened images
P–M, TCAD, CTAD, CTAD-TFT, CTD, and the proposed models. The in Sections 4.3 and 4.4.
images presented in row I of Figs. 11(a), 11(b), and 11(c) & 12(a),
12(b) and 12(c) are the noisy images of Fig. 10(a)–(f), respectively. 4.3. Weld defect segmentation using six sigma concepts
The corresponding parametric values given in Table 1 for various
models are used to smooth the input images. The corresponding On the smoothened image, a simple statistical method [9,28]
resultant images are given from second to seventh row in Fig. 11 is applied to find the control limits. The upper and lower con-
and Fig. 12. trol limits for gray value variations in the smoothened images are
The threshold parameter value (k) is adaptively taken in the μ S I + C σ S I and μ S I − C σ S I respectively, where μ S I and σ S I rep-
proposed model of the smoothened images in Figs. 11(a), 11(b), resent the mean and standard deviation of the pixel values of the
11(c), 12(a), 12(b), and 12(c) at row VII are k = 4, 5, 5, 5, 4, and smoothened image. C , a control constant, which is an integer in
5 respectively. Besides, the number of iterations t = 11, 19, 21, 20, the closed interval [1, 3]. During the image segmentation process,
12, and 16 respectively. It has been revealed that the threshold if a pixel value lies within the control limits, then it is considered
value of parameter k (obtained by the proposed adaptive selec- as the background pixel. Otherwise, it is a foreground pixel. In this
120 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
Fig. 12. Smoothened results; Column (a), (b) and (c) first row: Noisy weld X-radiography images of Fig. 10(d), 10(e), and 10(f), respectively; from row II to VII: corresponding
smoothened images obtained from P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed model, respectively.
way an image gets segmented into two homogeneous regions. In ity. For the original image as F (i , j ) and the denoised image as
order to identify the defective region more preciously during the G (i , j ) with size M × N, MSE can be defined as
segmentation, we considered 2-sigma standard by choosing the
M −1 N
−1
control constant C as 2. This segmentation procedure is followed 1 2
in all the experiments. Figs. 13–18 show the defect segmented re- MSE( F , G ) = F (i , j ) − G (i , j ) , (19)
MN
sults besides their 2D intensity profiles of a scan line that passes i =0 j =0
Fig. 13. Porosity weld defect segmentation and 2D intensity profiles of Fig. 11(a) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f) segmented results
and 2D intensity profiles of the line 75.
Fig. 14. Gas pore weld defect segmentation and 2D intensity profiles of Fig. 11(b) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f) segmented results
and 2D intensity profiles of the line 75.
122 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
Fig. 15. Undercut weld defect segmentation and 2D intensity profiles of Fig. 11(c) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f) segmented results
and 2D intensity profiles of the line 45.
Fig. 16. Slag inclusion weld defect segmentation and 2D intensity profiles of Fig. 12(a) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f) segmented
results and 2D intensity profiles of the line 45.
M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126 123
Fig. 17. Tungsten inclusion weld defect segmentation and 2D intensity profiles of Fig. 12(b) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f)
segmented results and 2D intensity profiles of the line 55.
Fig. 18. Crack weld defect segmentation and 2D intensity profiles of Fig. 12(c) using P–M, TCAD, CTAD, CTAD-TFT, CTD and the proposed models; (a)–(f) segmented results
and 2D intensity profiles of the line 85.
124 M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126
Table 2
Performance measure of the smoothened weld images.
and G, of common window size n × n. The common window size Furthermore, during the diffusion process, an adaptive threshold
is set as 3 × 3 in this study. A higher value of MSSIM indicates that parameter and automatic diffusion stopping criterion were intro-
the resultant image has more similarity to the original image. duced based on the Shannon entropy value that also decides the
effectiveness of denoising, while preserving the fine details. The
(2μ F μG + c 1 )(2σ F G + c 2 )
SSIM( F , G ) = (21) segmentation result of smoothened images show that the pro-
(μ2F + μ2G + c 1 )(σ F2 + σG2 + c 2 ) posed model can effectively remove the noises in noisy weld X-
radiography image. The denoising performance of the proposed
where μ F is the average of F ; μG is the average of G; σ F2 and σG2
model was evaluated using MSE, SNR, PSNR, Entropy and MSSIM
are the variance of F and G respectively; σ F G is the covariance;
metrics and the measurement results ensure that our model pro-
c 1 and c 2 are stabilizing parameters, generally c 1 = 0.05 and c 2 =
vides better results and outperforms those of others.
0.05. Recall that the covariance between F and G is defined as
L −1
1 Acknowledgments
σF G = ( F i − μ F )(G i − μG )
L−1
i =0 The authors would like to acknowledge the support rendered
Their mean SSIM is defined as by the managements of SASTRA University and Indira Gandhi Cen-
tre for Atomic Research, Kalpakkam by way of providing necessary
1
W
infrastructure for this research. This work was a part of generic
MSSIM( F , G ) = SSIM( F j , G j ) (22)
W image processing tool named “Bhadraloka” being developed for
j =1 the research project entitled “Development of techniques for pro-
where W is number of local windows of size 3 × 3 in the image. cessing radiographic images for automated detection of defects”
Table 2 shows the performance measure of the proposed (No. 2013/36/40-BRNS/2305) in Dot NET platform at SASTRA with
method with existing ones on smoothened images. The perfor- the funding assistance from Board of Research in Nuclear Sciences
mance of the proposed method ensures the accuracy of image (BRNS), Department of Atomic Energy, Government of India.
smoothening.
Appendix A. Supplementary material
5. Conclusions
Supplementary material related to this article can be found on-
In this paper, we have proposed a modified anisotropic diffu- line at http://dx.doi.org/10.1016/j.dsp.2017.05.014.
sion scheme based denoising, feature space gradient preserving
and weld defect detection for noisy weld X-radiography images. References
The noisy defects in a weld X-radiography image are unclear and
have different sizes. In order to detect the weld defects, the noisy [1] R. Ricardo, D. Mery, State-of-the-art of weld seam inspection by radiographic
image must be denoised or smoothened out and the fine details testing, part I: image processing, Mater. Eval. 65 (2007) 643–647.
must be preserved. The existing models have considered the pa- [2] J. Mohan, V. Krishnaveni, Y. Guo, A survey on the magnetic resonance image
denoising methods, Biomed. Signal Process. Control 9 (2014) 56–69.
rameters assigned randomly during diffusion process. Since the im-
[3] C. Pal, A. Chakrabarti, R. Ghosh, A brief survey of recent edge-preserving
age has noise, the defects remain unclear. Our diffusion model in- smoothing algorithms on digital images, Proc. Comput. Sci. (2015) 1–40.
corporates local probability value of gray levels into the traditional [4] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion,
diffusion coefficient function to denoising and defects preserving. IEEE Trans. Pattern Anal. Mach. Intell. 12 (1990) 629–639.
M. Malarvel et al. / Digital Signal Processing 68 (2017) 112–126 125
[5] E. Catte, P.I. Lions, J.M. Morel, T. Coll, Image selective smoothing and edge de- Gopalakrishnan Sethumadhavan received the
tection by nonlinear diffusion, SIAM J. Numer. Anal. 29 (1992) 182–193. M.Sc. and M.Phil. degrees in Mathematics from Bhara-
[6] L.I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal thidasan University, Tiruchirappalli, India, in 1985 and
algorithms, Physica D 60 (1992) 259–268. 1987 respectively and an M.Tech. degree in Quality
[7] D.M. Tsai, S.M. Chao, An anisotropic diffusion-based defect detection for sput- Assurance Technology from Regional Engineering Col-
tered surfaces with inhomogeneous textures, Image Vis. Comput. 23 (2005)
lege, Tiruchirappalli in 1995. He has also acquired a
325–338.
degree in Master of Computer Applications. He re-
[8] S.M. Chao, D.M. Tsai, Astronomical image restoration using an improved
anisotropic diffusion, Pattern Recognit. Lett. 27 (2006) 335–344. ceived the Ph.D. degree from SASTRA University, India,
[9] S.M. Chao, D.M. Tsai, An anisotropic diffusion-based defect detection for low- for his work in Biomedical Image Processing and Its
contrast glass substrates, Image Vis. Comput. 26 (2008) 187–200. Applications to Clinical Diagnosis in 2006. He is a member of Computer
[10] D.M. Tsai, C.C. Chang, S.M. Chao, Micro-crack inspection in heterogeneously Society of India. He is currently a Professor of Department of Computer
textured solar wafers using anisotropic diffusion, Image Vis. Comput. 28 (2010) Applications, School of Computing, SASTRA University. His research inter-
491–501. ests are in spatial domain image analysis with more emphasis on adaptive
[11] S.M. Chao, D.M. Tsai, Anisotropic diffusion with generalized diffusion coefficient segmentation and feature extraction. He also works in evolutionary algo-
function for defect detection in low-contrast surface images, Pattern Recognit. rithms. He is especially interested in developing software in medical and
43 (2010) 1917–1931.
industrial applications using image processing and soft computing tech-
[12] S.M. Chao, D.M. Tsai, An improved anisotropic diffusion model for detail- and
niques. He also bagged several best paper awards. He has over 30 years of
edge-preserving smoothing, Pattern Recognit. Lett. 31 (2010) 2012–2023.
teaching experience and 15 years of research experience.
[13] H. Tian, H. Cai, J.H. Lai, X. Xu, Effective image noise removal based on difference
eigenvalue, in: 18th IEEE Int. Conf. Image Process, 2011, pp. 3357–3360.
[14] V.B. Surya Prasath, D. Vorotnikov, Weighted and well-balanced anisotropic dif-
fusion scheme for image denoising and restoration, Nonlinear Anal., Real World
Appl. 17 (2014) 33–46. Purna Chandra Rao Bhagi joined Indira Gandhi
[15] C.A.Z. Barcelos, M. Boaventura, E.C. Silva, A well-balanced flow equation for Center for Atomic Research, Kalpakkam after Metallur-
noise removal and edge detection, IEEE Trans. Image Process. 12 (2003) gical Engineering bachelor’s degree from Andhra Uni-
751–763. versity, India, and obtained Ph.D. from Saarland Uni-
[16] H. Luo, L. Zhu, H. Ding, Coupled anisotropic diffusion for image selective versity, Germany. Currently, he heads Non-Destructive
smoothing, Signal Process. 86 (2006) 1728–1736. Evaluation (NDE) Division working towards develop-
[17] J. Yu, Y. Wang, Y. Shen, Noise reduction and edge detection via kernel ment of advanced techniques for detection and quan-
anisotropic diffusion, Pattern Recognit. Lett. 29 (2008) 1496–1503. titative evaluation of defects, residual stresses and mi-
[18] D. Fang, Z. Nanning, X. Jianru, Image smoothing and sharpening based on non-
crostructures by placing emphasis on sensors, numer-
linear diffusion equation, Signal Process. 88 (2008) 2850–2855.
ical modeling, image processing, quality assurance and automation. He is
[19] Q. Chen, P. Montesinos, Q. Sen Sun, D. Shen Xia, Ramp preserving Perona–Malik
responsible for development of comprehensive technologies for NDE of
model, Signal Process. 90 (2010) 1963–1975.
[20] Z. Guo, J. Sun, D. Zhang, B. Wu, Adaptive Perona–Malik model based on the
steam generator tubes, clad tubes and main vessels of fast reactors. Dr.
variable exponent for image denoising, IEEE Trans. Image Process. 21 (2012) B.P.C. Rao has a book and six book chapters and an Encyclopedia of Mate-
958–967. rials article to his credit. He is a recipient of Indian Nuclear Society Gold
[21] Y.Q. Wang, J. Guo, W. Chen, W. Zhang, Image denoising using modified Medal, National NDT Award and several best paper awards.
Perona–Malik model based on directional Laplacian, Signal Process. 93 (2013)
2548–2558.
[22] B.J. Maiseli, H. Gao, Robust edge detector based on anisotropic diffusion-driven
process, Inf. Process. Lett. 116 (2016) 373–378. Soumitra Kar received Bachelor’s degree in Elec-
[23] J. Xu, Y. Jia, Z. Shi, K. Pang, An improved anisotropic diffusion filter with semi- tronics and Telecommunication Engineering from Ja-
adaptive threshold for edge preservation, Signal Process. 119 (2016) 80–91. davpur University, Kolkata, with 1st class honors in
[24] I.B. Mhamed, S. Abid, F. Fnaiech, Weld defect detection using a modified 1985. He completed Ph.D. from CSE department of IIT
anisotropic model, EURASIP J. Adv. Signal Process. 46 (2012) 1–12. Bombay in 2002. He joined Bhabha Atomic Research
[25] S.A. Halim, A. Ibrahim, Y.H. Manurung, PDE-based model for weld defect detec-
Centre (BARC), Mumbai and received Homi Bhabha
tion on digital radiographic image, Int. J. Signal Process. Syst. 1 (2013) 146–151.
Award for securing 1st rank in Electronics discipline
[26] C. Dang, J. Gao, Z. Wang, F. Chen, Y. Xiao, Multi-step radiographic image en-
hancement conforming to weld defect segmentation, IET Image Process. 9
in the orientation course in Nuclear Engineering at
(2015) 943–950. BARC Training School (in 1986). Currently he is a sci-
[27] D. Mery, V. Riffo, U. Zscherpel, G. Mondragón, I. Lillo, I. Zuccar, H. Lobel, M. entist in grade H at BARC, Mumbai, and holds the position of Head of
Carrasco, GDXray – The Grima Database of X-Ray Images, Department of Com- Image Processing and Machine Vision Section in EISD division. He is also
puter Science, Universidad Católica de Chile, in Collaboration with Institute for a Professor, Engg. Sciences, at Homi Bhabha National Institute (HBNI),
Materials Research and Testing (BAM), Berlin, 2015. Mumbai. He has developed many techniques for face recognition, machine
[28] S. Gopalakrishnan, S. Srinivasan, Border detection and cancer propagation vision for automation, and image analysis for inspection. His research area
on spectral bands of malignant melanoma using six sigma threshold, in: of interest includes Image Processing, Artificial Vision and Biometrics. He
8th IEEE/ACIS Intl. Conf. on Computer and Information Science, ICIS, 2009,
has over 30 years of research experience and over 14 years of teaching
pp. 586–592.
experience.
[29] O. Zahran, H. Kasban, M. El-Kordy, F.E.A. El-Samie, Automatic weld defect iden-
tification from radiographic images, NDT E Int. 57 (2013) 26–35.
Arunmuthu Krishnan received the M.Sc. degree gramming and computer vision. He was one of the people responsible for
in Physics from the Madras University, India, in 2000. the development and realization of the X-ray computed tomography sys-
He has also acquired Master Diploma in Information tem at IGCAR. He selected as winner of The Ron Halmshaw Award 2010
Technology in 2001. He joined Indira Gandhi Centre for the best paper published in Insight on any aspect of industrial radio-
for Atomic Research as scientific Officer in 2003. His graphy or radiology. He is a member of Indian Society of Non-Destructive
research interests include development of X-ray com- Testing and Indian Institute of Metals.
puted tomography, image reconstruction using limited
projections, image and data processing, real-time pro-