Field of the invention:-
The invention relates to the field of continuous casting of steel in general and to
a method for online evaluation of steel slab quality by surface crack image
segmentation in particular. The evaluation is done on a continuous basis during
production of steel slabs during production in a continuous casting process.
Background and prior art:-
Within the continuous casting sub-process of the steel manufacturing process
where molten steel is passed through a water-cooled near-vertically-aligned
lubricated mold of rectangular cross-section about a meter long to emerge in the
form of a continuous strand consisting of a solidified shell encapsulating molten
material. This strand is further cooled to complete solidification using water
sprays even as the orientation is changed to horizontal using rollers, in the
process generating thermal and bending stresses and exacerbating cracks
originating in the mold-cooling stage, to the extent that the cracks become
visible to human observation and represent significant deterioration of steel
product quality.
Therefore, there is need for automatic inspection and classification of "slabs"-
obtained by cutting the strand section-wise at near-uniform lengths-based on
nature and intensity of these surface cracks so that the appropriate downstream
processing can be automatically categorized, and furthermore from a different
perspective, a relational mapping between the multiple continuous casting
parameters and the quality of resultant slab be determined.
Reference is made to patents US4561104, US4519041 and US306954 as well as
the following publications.

1. World Steel University website, continuous casting link:
http://www.steeluniversity.org/content/html/eng/default.asp?catid=27&pageid=
2081271519
2. J.K.Udupa and S.Samarasehara, "Fuzzey Connectedness and Object Definition
Theory, Algorithms and Applications in Image Segmentation" , Graphical Models
and Image Processing, Vol. 58, No.3, May 1996, pp.246-261.
3 P.K.Saha, J.K.Udupa, D.Odhner, "Scale-Based Fuzzy Connected Image
Segmentation: Theory, Algorithms, and Validation", Computer Vision and Image
Understanding, Vol 77, 2000, pp. 145-174.
4. J.K.Udupa, P.K.Saha, R.A.Lotufo, " Relative Fuzzy Connectedness and Object
Definition: Theory, Algorithms, and Applications in Image Segmentation", IEEE
Trans. On Pattern Analysis and Machine Intelligence, Vol. 24, No. 11, Nov. 2002,
pp. 1485-1500.
5. J.Canny, " a computational Approach to Edge Detection", IEEE Transactions
on Pattern Analysis and Machine Intelligence, Vol. 8, No. 5, Nov. 1986. pp. 679-
698.
6. J.R.Parker, http://paaes.cpsc.ucal garv.ca/~parker/501/edgedetect.pdf.
7. R.O.Duda and P.E.Hart," Use of the Hough Transformation to Direct Lines and
Curves in Pictures", Communications of the ACM, Vol. 15, Jan. 1972, pp. 11-15.
In the continuous casting process of steel manufacture [Ref 1], molten steel is
poured into a vertically aligned water-cooled mold of rectangular cross-section,

from where it emerges as a continuous 'strand'- a solidified shell encapsulating
liquid material. This is further cooled by water sprays and bent using rollers till
complete solidification and horizontal alignment is attained. The thermal and
bending stresses developed in the process sometimes exacerbate cracks that are
visible on strand surface. These cracks represent significant quality deterioration
of the final downstream product attained after rolling, chemical and heat
treatment of the discrete slab, the slab itself generated by gas-cutting the
solidified strand at regular intervals.
Efforts have been made for more than two decades to automatically recognize
and classify cracks on emergent strand surface in the steel continuous casting
process, see patents US4561104 and US4519041. Filtering out the crack/s from
clutter in the image, caused by scales, depressions and other deformities, is a
major problem that has prevented automatic slab quality inspection from gaining
wide acceptance. Most of these methods depend on edge detection algorithms
and intensity (grayscale) thresholding, combined with heuristic rules, to segment
cracks. In the present invention, the relative fuzzey connectedness technique
developed by Udupa and others Reference[l], [2], [3], [4], has been used in
conjunction with Canny edge detection technique reference [5] and the Hough
transform based line segmentation technique (Patent US3069654) to segment
cracks on strand surfaces. This invention is successful in recognizing, delineating
and classifying cracks and isolating it from clutter in an entirely automated and
real time setting, as results show. The present invention seeks to overcome the
above drawbacks of prior art.
Obfects of the invention:-
An object of the invention is to provide a fully automated, real time and reliable
method of classification of steel slabs during their production by continuous

casting.
Another objects of the invention is to eliminate human inspection of the cracks
on such steel slabs.
A further object of the invention is to arrive at true segmentation of the cracks
from image of slab surface for accurate image recognition.
A still further object of the invention is to provide a method that relates to the
real-time conditions of this multi-parametric process to determine the emergent
strand quality.
Description of the invention:-
An important step in the steel manufacturing process is the conversion of liquid
steel to solid slabs by the continuous casting method. These slabs are further
rolled into thin sheets according to customer requirements. Quality issues that
may arise in the solidification process have significant downstream impact, and
mitigation of these issues is a major technological challenge. One aspect of
quality deterioration is the formation of surface cracks on solidifying slabs. A
novel method is described here to automatically recognize and classify these
cracks according to intensity and type in real time, from slab surface image
captured by cameras installed near the slab production line. Cracks thus
identified directly contribute to the classification of generated slabs according to
quality and thus determine their downstream processing. Equally importantly,
this quality identification approach is being used to refine and verify a process
expect system that will directly predict slab quality from real time conditions of
the multi-parametric continuous casting process, and thus refine the process
itself in response to the above-mentioned technological challenge.

The method for automated crack identification accordance with the invention is
based on the method of relative fuzzy connectedness, wherein 'seed points'
(seed pixels) are identified within the image under analysis on objects of interest
and the background, and these and the background, and these seed points
compete among themselves to attract all other pixels and their success depends
on the relative strength of fuzzy connectedness of the pixel to the selected seed
points, and where the 'strength of fuzzy connectedness' is obtained from affinity
relations defined on considerations of distance, homogeneity and feature
similarity between pixel and seed points, wherein the process of identification of
seed points itself is by manual inspection and hence not automatic.
Elimination of cracks and other quality deteriorations during solidification is a
major technological challenge for the continuous casting process. In response,
efforts are underway by the applicant to develop a caster expert system that will
relate the real-time conditions of this multi-parametric process to the emergent
strand quality, and consequently improve and refine the process itself. To aid in
this effort, a mechanism is needed for real-time assessment of the strand quality,
and this work is in that direction. It is a technique for real time recognition and
delineation of racks from images of strand surfaces captured by cameras
installed along the emergent strand path. Apart from aiding in the effort towards
a caster expert system, this technique facilitates automated classification of
generated slabs based on quality for differentiated downstream processing-a
task performed manually till now.
For crack identification a sequence of steps is followed, first using Canny edge
detection technique to locate all possible edge in the image, and second using
the Hough transform to identify near-straight lines among these edges. The third
and critical step uses the Relative Fuzzy Connectedness approach to segment
cracks that tend to be aligned along these straight lines. Once a crack is

identified, its intensity and type can be extracted using simple rules. In the crack
recognition task, a major problem is the isolation of cracks from clutter caused
by scaling, depressions and other deformities. As the extent of clutter varies
significantly across images, reliable automatic filtration becomes unattainable
using crisp threshold-dependent methods. The fuzzy framework in which the
relative connectedness technique works is able to handle this problem with ease,
as results show. The only manual insertion in this technique is the selection of
object and background seed points for process initiation; Canny and Hough
sequential pre-processing is used to identify these points, in the process
generating a method for slab surface quality evaluation that is fully automatic,
reliable and real time.
The fuzzy connectedness technique of the present invention addresses two
aspects in an image, namely graded composition (variation in intensity) and
hanging-togetherness (connectedness), within a fuzzy framework. Any crisp, i.e.
non-fuzzy, approach is compelled to define thresholds on both these aspects,
which enforce a level of unnaturalness, or distinctness from the approach of the
human mind, in object identification. This is less desirable from the viewpoint of
automation of object recognition, and possibly explains why earlier methods for
surface cracks automated recognition have not been successful. The way relative
fuzzy connectedness works is that, first, seed points are identified on objects of
interest and background, second, the seed points compete among themselves to
attract all other pixels and their success depends on the relative strength of fuzzy
connectedness of the pixel to the selected seed points, where the 'strength of
fuzzy connectedness' is obtained from affinity relations defined on considerations
of distance, homogeneity and similarity to seed points. At the end of this process
the objects and background are clearly delineated.

Selection of seed points is a critical step in the object identification process and
unfortunately is the only non-automated step. Here a coupling of Canny and
Hough techniques is used to define the seed points. Results presented in this
specification show that the method works successfully, within time cycles that
are slightly larger that real-time requirements, which the applicants believe can
be significantly cut down in future.
Brief Pescription of the Accompanying Drawings:
Figure 1 is a schematic diagram of continuous casting configuration.
Figure 2 is a schematic diagram of the emergent steel strand, image capture
and transfer across LAN to remote PC for image processing according
to the invention.
Figure 3 is a schematic diagram of crack identification software modular inter-
relations according to the invention.
Figure 4a shows original slab image captured on camera, see the long crack.
Figure 4b shows after performing edge detection using Canny's method.
Figure 4c shows detected line superposed on original image after segmenting
lines on Canny edge image Hough transform.
Figure 4d shows narrow window window around identified line selected for
analysis; some clutter from scale/deformities is still visible.
Figure 5a shows original slab image captured on camera, showing the shorter
crack and extensive clutter.
Figure 5b shows profusion of edge after performing edge detection using
Canny's method.
Figure 5c shows detected line superposed on original image after segmenting
lines on Canny edge image using Hough transform.
Figure 5d shows detected line superposed on original image still visible after
segmenting lines on Canny edge image using Hough transform.

A exemplary embodiment of the invention will now be described in detail with
the help of the accompanying drawings. However, there can be other
embodiments of the invention, which are deemed covered by the present
description.
Description Of The Preferred Embodiment:-
A CCD camera with an infrared filter is placed at an angle between 40° and 50°,
preferably at 45° from the vertical symmertric plane to the slowly moving strand,
at a distance of around 8 meters from it, as schematically shown in Fig. 2. The
strand moves at approximately 1.5 meters/min, and every minute an image is
transferred across local area network (LAN) to the remote PC where image
processing is performed. The image processing software executes a sequence of
operations on the gray-scale image, first, detection of edges using Canny
method, second, identification of straight line/s among these edges using the
Hough transform, third, contrast-stretching of the image, fourth, delineation of
cracks using method of relative fuzzy connectedness, and last in Summarizer,
see Fig. 3, classification of identified cracks based on length and intensity using a
simple set of rules.
Canny edge detection technique is one among several edge detection methods
that work on the gradients between neighboring pixel intensities in an image.
The Canny method finds edges by looking for local maxima of the gradient of an
image. The gradient is calculated using the derivative of a Gaussian filter. The
method uses two thresholds, to detect strong and weak edges, and includes the
weak edge in the output only if they are connected to strong edges. The Hough
transform (Patent US3069654, also reference [7]) maps edge points in the image
original (x,y) space to curves in the (r, 0) space, where r and 6 represent
distance of the line from an origin point in the (x,y) space, and the inclination of
normal drawn from origin to that line (inversely, a point in the (r,9) space maps
onto a line in the (x,y) space). Collinear points in the (x,y) space represent

intersecting curves in the (r,Ө) space, and the points in the (r,Ө) space that have
maximum number of curve intersections can be considered to represent straight
edge in the (x,y) space. This method is used to identify straight lines among the
edges identified by Canny method in the image. Figures 2(a-c) illustrate the
results of these operations on a test image. At this stage if no lines are identified,
it is assumed that no cracks exist in the image under consideration, and
processing for the image is aborted in the program, which then waits for the
next image.
When the above process identifies a crack line in the image, the next stage of
processing- delineation of cracks using the method of relative fuzzy
connectedness- starts. The relative fuzzy connectedness technique identifies
each pixel in the selected image as belonging to either object or background,
based on the relative strength of connectivity between the selected pixel and the
respective seed pixels. The strength of connectivity between any two pixels in an
image is defined as the strength of the strongest of all possible paths between
the two pixels lying entirely within the image. The strength of a selected path
between two pixels is defined as the weakest affinity between two consecutive
pixels lying on the path anywhere along its length, analogous to the strength of
a chain being identified on its weakest link. All the above concepts are defined in
a fuzzy framework, i.e. the degree of affinity varies continuously from 0 to 1, the
strength of connectivity varies likewise, and the degree of belonging of a pixel to
object and background also varies from 0 to 1. Thus a pixel at any location
belongs both to object and to background; and in a certain subspace of the
image the degree of belonging of pixels to object becomes higher.
The relative fuzzy connectedness method is used to delineate crack objects.
From the set of pixels that lie on the Hough transform-identified line on the

image, the darkest two are selected as seed points, provided they are not very
close to either boundary. Two to four points of intermediate brightness are
selected as background seed points, note that extremely bright points are not
selected as background since these tend to 'push' clutter into the crack object. In
the actual usage of relative fuzzy connectedness, the mathematical relations for
affinity are strictly followed, while the algorithms for evaluation of relative
strength of connectivity and thus delineation of crack objects are generated
afresh and distinct from those presented in the prior art. Once the crack is
identified, the average grayscale value or 'darkness' of the pixels that lie within
the crack is evaluated, as also the length of the crack. Together, they indicate
the intensity of the delineated crack.
Two crack images and their analysis is presented for perusal, the first
representing a good crack from the perspective of image segmentation, and the
second representing a poor one.
Figure 4a shows an image of a good crack- long and deep, with average clutter
in the background. The edges identified by Canny method are shown in Fig. 4b.
Edges corresponding to the long crack are clearly seen. Figure 4c shows the line
obtained after Hough transform. It is superimposed on the original image to
demonstrate correctness of location. Note that it is just a complete line, telling
that a crack lies somewhere along its length, but nothing about the crack itself.
The first point of observation about Fig. 4d is the width of the image; only a
narrow section around the identified line has been taken. The second is that the
figure is not an image; it represents either the object, or the background in
white. The third point is that the crack has been nicely delineated with minimal
clutter.

Figure 5a shows an image of a crack that is difficult to segment, due to tine
presence of intense clutter and the relative weakness of the crack. Figure 5b
shows the Canny edges- note their multiplicity in the midst of which the straight
edges related to crack seem to be lost. Figure 5c shows that the Hough
transform has successfully picked up the crack line. And Fig. 5d illustrates that
the crack has again been successfully delineated, as comparing with the original
image will show, though some clutter has also been picked up-which should not
affect the intensity identification procedure as described.

We Claim:
1. A method for online evaluation of steel slab quality in production phase by
surface crack image segmentation comprising the steps of:
- classification of the cracks according to intensity and type in real time from slab
surface images captured by CCD cameras installed near the slab production line;
- identification of a crack by Canny edge detection technique to locate all
possible edge in the image;
- extracting intensity of the crack;
- isolation of a crack from a clutter caused by scaling depressions and other
deformities with the help of fuzzy logic and
- selection of object and background seed points for process initiation;
characterized in that the method evaluates slab surface quality automatically,
reliably and in real time.

2. A method as claimed in claim 1, wherein said slab surface images captured by
CCD camera are transferred across LAN to remote PC for image processing.
3. A method as claimed in claims 1 and 2, wherein said image processing
comprises the steps of detection of edge by Canny method, identification of
straight lines among these edge by Hough transform, contrast-sketching of the
images, isolation and delineation of the cracks by relative fuzzy connectedness
and classification of identified cracks based on length and intensity.
4. A method as claimed in claims 1, 2 and 3, wherein said relative fuzzy
connectedness identifies each pixel in the selected image as belonging to either
object or background, based on the relative strength of connectivity between the
selected pixel and the respective seed pixels.

5. A method as claimed in claim 1, wherein the CCD camera is provided with an
infra-red filter, is placed near the moving strand at an angle between 40° and
50° most preferably at 45° from the vertical plane and captures an image of the
moving strand once every minute.

A method for online evaluation of steel slab quality in production phase by
surface crack image segmentation comprising the steps of:
- classification of the cracks according to intensity and type in real time from slab
surface images captured by CCD cameras installed near the slab production line;
- identification of a crack by Canny edge detection technique to locate all
possible edge in the image;
- extracting intensity of the crack;
- isolation of a crack from a clutter caused by scaling depressions and other
deformities with the help of fuzzy logic and
selection of object and background seed points for process initiation;
characterized in that the method evaluates slab surface quality automatically,
reliably and in real time.