FIELD OF THE INVENTION
The invention generally relates to a Real Time Radioscopy system, which,
automatically indicates Acceptance/Rejection of Straight Tube Butt weld joints
based on the weld-quality in accordance with the acceptance standard. More
particularly, the invention relates to a Real Time Radioscopy System for
automatic detection of defects in weld joints of Straight Tube Butt Welds.
BACKGROUND OF THE INVENTION
Conventionally, evaluation of the result of Real Time Radioscopy of the Straight
Tube Butt weld is carried-out manually by qualified and experienced operators.
However, manual evaluation involves certain limitations like subjectively, fatigue
etc, which would affect the reliability and productivity.
Automatic Defect Recognition (ADR) is known in the art, and ADR systems are
available for generic Applications, on Non-Destructive Testing, and specialized
applications like Castings, Machined components etc.
US Patent No.4896278 entitled "Automatic Defect Recognition system" is
applicable for general NDT applications of various products.
US Patent Publication No. 20100215150 provides a real-time method for
navigation inside a region of interest, for use in a radiography unit including an
X-ray source with an X-Ray detector facing the source (cradle), and a support
(table) on which an object to be radiographed, containing the region of interest,
can be positioned.

Indian Patent Application NO.2207/KOL/08 of 23rd December 2008 filed in the
name of Bharat Heavy Electricals Limited relates to Automatic Defect Recognition
system for Real Time Radioscopy of Hancock Valves. According to the disclosure
of prior art, the process defect classification is done after defect detection only.
OBJECTS OF THE INVENTION DISCLOSED
It is therefore an object of the invention to propose an automatic defect
recognition system in a real time radioscopy of straight-tube butt weld joint,
without the limitations of the prior art.
Another object of the invention is to propose an automatic defect recognition
System in a real time radioscopy of straight-tube butt weld joint, which is
capable to analyze and classify the defects based on the size, shape, orientation
and location of the detected defect(s).
Still another object of the invention is to propose an automatic defect recognition
System in a real time radioscopy of straight-tube butt weld joint, which is
enabled to indicate acceptance or Rejection of the Weld joint based on
comparison of the real-time data with the acceptance standard for Straight Tube
Butt Weld joints.
Yet another object of the invention is to propose an automatic defect recognition
system in a real time radioscopy of straight-tube butt weld joint which replaces
the prior art evaluation of Straight Tube Butt Weld joints by human operators
with automatic evaluation.
A further object of the invention is to propose an automatic defect recognition
system in a real time radioscopy of straight-tube butt weld joint which enhances

the speed and efficiency of the evaluation process by using artificial neural
networking techniques.
SUMMARY OF THE INVENTION
Digital X-Ray images are the two dimensional representation of the internal
structure of the exposed weld joint, in terms of the variation in radiation
absorbed and transmitted by each and every point in the exposed region of the
weld joint. In an RTR image, as represented by a display device for being
evaluated, different points (which are called pixels, in the case of a Digital image,
which is used in the case of this invention) are represented by corresponding
Gray value, with Black and White as extreme values (e.g. For an 8-bit image,
there are 256 gray value levels, with 0 corresponds to pure black and 255
corresponds to pure white, and other intensity levels between these two).
During evaluation of the digital image, the decision of acceptance or Rejection is
made based on the deviation of an image in terms of the Gray value pattern,
from the image pattern of a good joint. The extent of this deviation, which is
quantified as the defect details, is accessed and is compared with the Acceptance
standards, which specify the acceptance limits, in order to take the decision of
Acceptance/ Rejection. Further, evaluating the image pattern with the similar
patterns, increases reliability as well as speed of the evaluation, and enhances
expertise of the operators.
According to the invention, Digital X-Ray image of the Straight Tue Butt Weld,
are captured by the imaging device and a pixel analysis is applied for detection
of different classes of potential defects in the weld. Depending upon the outcome
of the analysis for each class of defects, the detection results are displayed. In

some cases, where the image quality is not adequate enough to recognize the
defect with certainty, the system calls for operator intervention.
The present invention is different from the prior art in Two aspects, viz, it's of
specified Nature as it is specific to Digital Radiography with Digital Flat Panel, for
straight tube butt (STB) weld joints, and the property, which enables it to
improve the accuracy, and speed of defect recognition through it's learning by
artificial neural network techniques.
The disclosed invention unlike the prior art does the Automatic evaluation of the
weld joint based on its radiographic images.
Thus, the ADR system of the invention is for the defect recognition and
characterization of Straight Tube Butt (STB) Weld joints based on their Real Time
Radioscopic images. The approach used by this system is different from the prior
art for the features that STB welds detection procedure adapted for different
defects are of different type and hence the defect detection and classification
occurs together.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS
Figure 1 - shows a pictorial representation of defect recognition stages;
Figure 2 - shows an input Image to the ADR in DICONDE format having a low
contrast;
Figure 3 - shows the contrast-enhanced input image of Figure 2;
Figure 4 - shows the Region of Interest (ROI) extracted from the contrast-
enhanced image of Figure 3
Figure 5 - shows a sample image of Incomplete Penetration (ICP)
Figure 6 - shows a sample image of Gas Hole

Figure 7 - shows a sample image of Burn-Through
Figure 8 - shows the basis for ICP feature extraction
Figure 9 - shows the contour of a Burn-Through image
Table 1 shows the network results.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT OF THE
INVENTION
According to the invention, a System consists of a constant potential mini-focal
X-Ray equipment provided with a swiveling device, an Amorphous Silicon Digital
Flat Panel (Digital Detector Array) acting as the Detector / Imaging device
including means for Image acquisition and processing of the Digital X-Ray
images and means for Defect recognition and Evaluation of Straight Tube Butt
weld joint.
When the joints to be automatically evaluated are exposed to the X-Ray, the
image is captured by the Imaging device and obtained in the Display device for
Image Acquisition. The same is processed and the Image quality is enhanced by
the means for Image Processing associated with the Digital Flat Panel.
The image acquired from the Digital Flat Panel is of low contrast and therefore
has very low signal-to-noise ratio. To make the image more suitable for further
processing, it is necessary to perform contrast-enhancement. This enhancement
is done by reshaping the histogram of the image such that low valued pixels are
saturated and the other pixels are evenly distributed over the whole range. The
result of this operation can be seen in Figure 3.

The next operation is edge detection, to isolate the tube in the image. Once the
edges of the tube are detected, the orientation of the tube within the image is
detected by least-square fitting through the edges. Using this measure of
orientation, the tube is exactly extracted from the image.
The region of Interest (ROI) in the image is the area where defects are expected
to occur and hence reviewed. In case of an image of straight tube-butt welds,
the ROI includes the elliptical weld region and adjacent raw material region of
joining tubes. The next step is the extraction of the ROI, by determining it's
boundaries through analysis of the vertical summation of the extracted tube for
regions of heightened activity. This is done by applying a threshold of the result
obtained from normalizing, smoothening and then differentiating the summation
twice to extract only regions where a drastic change is present. The boundaries
of the ROI are extracted by scanning the result of the above described step.
The next operation is feature extraction. Since each defect manifests in a
different manner, different feature vectors are used for each defect. This process
uses Artificial Neural Network Training technique using Radial Bias function
(RBF). Each defect has its own detection RBF which is trained on the feature
vector specific to the defect.
There are basically three categories in which the defects are classified. One set
covers all the defects like Gas Holes, Porosities and other rounded indications.
The second type includes incomplete Penetration (ICP) and Lack of Fusion. And
the third type includes Burn through and Excess Penetration. Gas Holes are seen
as prominent white dots in the ROI. The ROI is first median filtered (with a
window of 20) and then subtracted from itself. This removes the larger features
from the image, while emphasizing the smaller defects. This image is again

median filtered with a kernel size of 5, which eliminates noise. The resulting
image is converted to binary and features are extracted from each white region
in it. The extracted features are: aspect ratio, area, length, breath and
roundness. These features are then passed individually to the RBF network for
classification. If even one gas hole is detected, the output of this block is the
RBF's output for that gas hole. If no gas hole is detected, the output of this block
is the mean of the RBF's output for all the features. If multiple gas holes are
detected, the output of this block is the mean of the RBF's output value for all
the gas holes.
For Incomplete Penetration, the most prominent indication is a breakage in the
root ellipse. The ROI is first median filtered with a window of 20 and then
subtracted from itself. This serves to emphasize the root in the ROI while
removing all other features. This image is then median filtered with a window
size of 5, and then converted to a binary image with a threshold of 0.008. This
image is split into two halves to separate the two sides of the root ellipse. The
vertical summation of the two halves is then calculated. Each of these
summations is used as a separate feature vector for the subsequent RBF
network. From the output of the network, it is possible to pinpoint the side in
which the ICP is present. The feature vector is able to represent the breakage in
the root ellipse that is characteristic of ICP and is of size 50. Since the
morphological difference between ICP and non-defective images is often subtle,
and the quality of the input images is poor, in some cases there is ambiguity
regarding the classification of images as ICP or non-defective. In such cases, an
exception is raised to the operator and the final decision is passed to him. This
ensures that the risk of false negatives is mitigated, as well as weeding out false
positives from the final classification. If the output of the RBF falls within a range
of values, an exception is raised to the operator. This range was determined by

experimentation, and is a trade-off between ICP detection accuracy and
minimizing the number of false positives.
Burn-Through is a defect that is caused by the complete destruction of the weld
region. It is manifested as a big, bright hole occasionally surrounded by small
black dots. The morphology of a burn-through is nearly impossible to specify and
characterize due to the variety of ways in which it can manifest. The feature
used to detect the burn-through is a vector formed by concatenating the vertical
and horizontal summation of the contour of the ROI. The RBF network is trained
on this vector.
The final decision regarding the classification is made by the final RBF, which
takes the output of the 3 previous RBF's as inputs.
A radial basis function (RBF) is a real-valued function whole value depends only
on the distance from the origin. The weighted sum of multiple radial basis
functions can be used to perform function approximation. A radial basis function
network is an artificial neural network that uses radial basis functions as
activation functions. It is a linear combination of radial basis functions. The RBF
that has been used in designing the network is Gaussian RBF. All the networks
were designed during the training process by iterating the network size to
achieve minimum error over the training vector.
Each RBF is trained using a batch process and then evaluated on a test set. The
whole assembly is evaluated on a test set comprised of images of all classes.

WE CLAIM:
1. An automatic defect recognition system in real time radioscopy of
straight tube butt welds, comprising:
a constant potential mini-focal X-Ray equipment provided with a
swiveling device;
a digital flat panel having means for acquiring and processing
digital X-Ray images;
a display device to exhibit the captured and processed images;
means for defect recognition and evaluation of straight tube butt
joints;
the system is configured to :
expose the butt joints to the X-Ray;
acquiring the X-Ray images of the butt joints and displaying in the
display device;
process the acquired images in respect of enhancing the image
quality;
detect the defects in the joints based on the processed data;
classify the detected defects into respective categories;

compare the defects with pre-stored values according to the
standard; and
automatically indicate acceptance or rejection of the defects.
2. The system as claimed in claim 1, wherein the processing of the
captured image data comprises reshaping the histogram of the image
such that: the low value pixels are saturated down, high valued pixels are
saturated up, while the mid-range pixels are evenly distributed over that
range of pixel intensity.
3. The system as claimed in claim 1 or 2, wherein the straight-tube image
is extracted by edge detection and measuring orientation of the tube.
4. The system as claimed in claim 3, wherein the region of interest ROI in
respect of the possible defects is extracted by determining the ROI -
boundaries through analysis of vertical summation of the extracted tube
image for regions of heightened activity.
5. The system as claimed in any of the preceding claims, wherein different
feature vectors are used for feature extraction in respect of different types
of defects, and wherein an artificial neural networking technique using
radial bias function (RBF) is used.
6. The system as claimed in claim 1, wherein the system is enabled to
classify the defects into at least three categories for example, gas holes,
porosities, other rounded indications; incomplete penetration, and lack of
fusion; and burn through and excess penetration.

7. The system as claimed in claim 5 or 6, wherein different neural networks
are used to detect different types of defects.
8. The system as claimed in claim 7, wherein the final decision on
acceptance is automatically displayed based on the result of the final
neural network which takes the output of all the previous neural networks,
which detect specific defects, as inputs.

The invention relates to an automatic defect recognition system for real time
radioscopy (RTR) of straight tube butt welds, comprising a constant potential
mini-focal X-Ray equipment, a swiveling device, a digital flat panel detector
(imaging device) capable of acquiring and processing digital X-ray images, a
display device to exhibit the captured and processed images, a means for defect
recognition and evaluation of straight tube butt weld joints. The RTR system is
configured to expose the butt joints to X-Rays, acquiring the X-Ray images of the
butt joints and displaying in the display device, process the acquired images in
respect of enhancing the image quality, detect the possible defects in the joints
based on the processed data, classify the detected defects into respective
categories, compare the defects with pre-stored values according to the standard
and automatically indicate acceptance or rejection of the defects.