Illumination-robust Recognition and Inspection in Carbide Insert Production

ACEEE Int. J. on Electrical and Power Engineering, Vol. 02, No. 03, Nov 2011

Illumination-robust Recognition and Inspection in
Carbide Insert Production
R. Schmitt1 and Y. Cai1
1
Laboratory for Machine Tools and Production Engineering, RWTH Aachen University, Aachen, Germany
Email: {R.Schmitt, Y.Cai}@wzl.rwth-aachen.de

Abstract—In processes of the production chain of carbide processing methods [2] [3]. The chip-former geometry turns
inserts, such as unloading or packaging, the conformity test out to be the most distinctive feature and a significant quality
of the insert type is performed manually, which causes a characteristic of carbide inserts. Compared to the other three
statistic increase of errors due to monotony and fatigue of quality features the chip-former geometry is the single feature,
workers as well as the wide variety of insert types. A measuring which represents an insert uniquely. In this paper, a measuring
method is introduced that automatically inspects the chip-
method is proposed for inspection and classification of the
former geometry, the most significant quality feature of
inserts. The proposed recognition approach is based on local chip-former geometry. This technique realises a robust and
energy model of feature perception and concatenates the phase automated measurement and inspection of carbide insert in
congruency in terms of local filter orientations into a compact the production line.
spatial histogram. This method has been tested on several
inserts of different types. Test results show that prevalent II. RELATED WORK
insert types can be inspected and classified robustly under
illumination variations. A. Object Recognition
Object recognition in computer vision is the task of
Index Terms—optical measurement, industrial image finding given objects in an image or video sequence and
processing, testing and inspection, local energy, phase classifying them into the known object types. The central
congruency, feature description, Gabor wavelets, illumination
problem of object recognition is how the regularities of objects
invariance
are extracted and recognized from their images, taken under
different lighting conditions and from various perspectives
I. INTRODUCTION
[4] [5] [6]. To capture the object characteristics, early
The product spectrum of cutting material manufacturers recognition algorithms are focused on using geometric models
consists of a large variety of insert geometries, which only of objects [7]. While the 3D model of the object being
differ by small details such as coating colour, edge radius, recognized is available, the 2D representation of the structure
plate shape and chip-former geometry. An example of a carbide of an object in image is compared with the 2D projection of
insert and its four remarkable features are shown in Fig. 1. the geometric model [7]. Other geometry-based methods try
Due to the extensive preparations and the long processing to extract geometric primitives (e.g., lines, circles, etc.) that
time, different production lots of inserts are combined for an are invariant to viewpoint change [8]. However, it has been
economic and reliable coating. Currently manual conformity shown that such primitives can only be reliably extracted
test regarding the plate type is used for the unloading of the under limited conditions (controlled variation in lighting and
coating machine. It is cost-intensive and involves a risk of viewpoint with certain occlusion) [7]. In contrast to early
interchanging for the subsequent packaging. Depending on geometry-based object recognition works, most recent efforts
the sorting tasks the error rate of manual insert sorting could are concentrated on appearance-based techniques [9]. The
be 5 to 35 percent [1] [2]. Generally machining with a false underlying idea of this approach is to extract features, as the
carbide insert type causes dimension faults in mechanical representations of object, and compare them to the features
work pieces. For this reason, this series process could be in the object type database, while 2D representations of
negatively affected and it causes unnecessary costs. objects viewed from different angles and distances are
available. Especially the local descriptors [10] [11] [12] do
not require object segmentation and describe and represent
objects with features directly extracted from local image
regions. Local descriptors can be computed efficiently, are
resistant to partial occlusion, and are relatively insensitive to
changes in viewpoint [13].
B. Local Feature Descriptor
There are two considerations to using local descriptors.
Figure 1. A carbide insert and its four features First, the interest points are localized in image. The interest
The insert quality features, coating colour, edge radius and points are often occurred at intensity discontinuity and
plate shape, could be inspected easily with simple image should be stable over transformations [14]. Second, a

© 2011 ACEEE 36
DOI: 01.IJEPE.02.03.92_14


description of the image region around the interest point is The developed prototype consists of the following hardware
built. To differentiate interest points reliably from each other, modules: illumination unit, camera/optic-system and
the feature description should ideally be distinctive, concise mechanical system. To record all necessary information for
and invariant over transformations caused by changes in the measurement and classification task, the illumination
acquisition pose and lighting conditions. While the concept of the insert inspection system employs a
localization of interest points and feature description are often combination of three different lighting types (Fig. 2). The
designed together, the solutions to these two steps are diffuse LED ring light as bright-field illumination provides a
independent [13]. One of the most widely used local feature homogeneous illumination of the entire insert without
representation is the scale-invariant feature transform (SIFT) disturbing reflections on the surface. By means of an LED
descriptor [11]. The SIFT uses interest points localized at ring light as dark-field illumination, a significant highlighting
local peaks in scale-space created with a pyramid of difference of chip-former geometry is enabled. The use of an LED-based
of Gaussian filters. At each interest point, the local image transmitted light illumination ensures a clear outer contour
gradients are sampled and weighted by a Gaussian, and then of the insert. To simulate an inline inspection of carbide inserts
represented in orientation histograms. The descriptor is in the production line, a practically relevant mechanical system
formed from a vector containing the values of all the is designed and implemented (Fig. 3). The camera/optic-
orientation histogram entries. The classification of objects is system and the insert handling system are synchronised by
performed by first matching each interest point in test image a trigger, which is associated with a motorized rotating disc.
independently to the database extracted from template As the insert comes into the camera view field, the reflexion
images, and then classifying the objects according to the light sensor releases the image acquisition. As a result, a
matches of interest points. Similar to the SIFT descriptor a well-focused insert image is provided.
descriptor is developed by applying principal components
analysis (PCA) to the normalized image gradient patch. The IV. CLASSIFICATION OF INSERT GEOMETRY
dimension of the feature descriptor is more compact than the
As shown in Fig. 1, the chip-former geometry appears
standard SIFT representation [12]. Another variation of the
exactly at the image edges of the insert surface. Based on
SIFT method is GLOH (Gradient Location and Orientation
this fact two methods (template matching with Hausdorff
Histogram). It considers more spatial regions for the
distance [20] and the geometric moments [21]), which classify
histograms. Through PCA the higher dimensionality of the
shapes with features extracted from the image edges, are
descriptor is reduced [13].
tested for the chip-former geometry recognition.
C. Work Piece Recognition
There are many approaches and methods to resolve the
work piece recognition problem [15] [16]. The methods based
on local feature descriptors are often applied to 2D images of
work pieces. SIFT characteristics are used as matching
features in [17]. By using an incline distance as the similarity
metrics of image matching, and setting a threshold to delete
the false matching points, the matching speed could be
improved. A descriptive vector, which consists of object
geometric properties, e. g. centroid, orientation angle and
height of objects, is produced in [18] to recognize assembly
parts. In [19] two image recognition systems based on random
local descriptors are introduced. The idea of random Figure 3. Machine vision prototype for insert inspection
descriptors of the image is that each neuron of the associative The test result of example insert sets demonstrates that it is
layer is connected to some randomly selected points on the not possible to robustly classify the insert sets with extracted
input image and such neuron its function using the brightness chip-former geometries, because these methods normally use
of these points. gradient-based edge detection operations (such as Canny
edge detector [21]) and the detected edges deviate
III. MACHINE VISION PROTOTYPE considerably depending on illumination variations and noise
[3]. The experiments with local feature descriptors such as
SIFT method [11] and PCA-SIFT [12] confirmed that their
feature representations are partially dependent on illumination
changes [3]. In contrast to the gradient-based methods, phase
congruency is a quantity that is invariant to changes in image
brightness or contrast [22]. Furthermore, phase congruency
is a dimensionless measure, whose value varies from a
maximum of 1 (indicating a very significant feature) down to
Figure 2. Developed Illumination unit 0 (indicating no significance). This absolute measure of the
© 2011 ACEEE 37
DOI: 01.IJEPE.02.03.92_14


significance of feature could be universally applied to any poor localization [23]. To compensate these effects, this
image irrespective of illumination [22]. The application of local measure is extended by consisting of sine of the phase
feature descriptor which is based on phase congruency of deviation, including a proper weighing of the frequency
image perception to face recognition shows this method is spread W (x) and also a noise cancellation factor T . This
insensitive to illumination and other common variations in
facial appearance [23] [24]. For an illumination-robust
modified measure, as given in (5), is more sensitive and yields
classification of chip-former geometry of the carbide insert, good localization of blurred features [22]
an approach is developed based on phase congruency. After
the phase congruencies of image are calculated (A. Phase PC ( x) 
 W ( x) A ( x) ( x)  T 
n n n
(5)
congruence), a feature histogram is formed, which builds a  A ( x)  
n n
distinctive and concise description of insert image (B. Feature
where  is a small positive constant and the cosine term(6)
in
histogram). Based on these extracted feature descriptors, the
test insert is identified with an object matching method (C. (1) is expanded by
Matching).
A. Phase Congruence
The local energy model, which has been developed in
[25], postulates that significant image features are perceived
at points where the Fourier components of the image are
maximally in phase. This model defines a phase congruency
function in terms of the Fourier series expansion of a signal
at location x as

PC ( x) 
 n
An ( x) cos(n ( x)   ( x ))
, (1)
 n
An ( x )

where An and n represent respectively the amplitude and Figure 4. Local energy and phase congruency from signal Fourier
components
local phase of the n -th Fourier component at position x .
To obtain frequency information (magnitude An and phase
 (x ) is the amplitude weighted mean local phase angle of
all the Fourier terms at the point n ) local to a point in a signal, logarithmic Gabor wavelet
filters are used [27]. This linear phase filters are in symmetric/
 ( x) 
 A ( x) ( x)
n n n
antisymmetric quadrature pairs and preserve phase
 A ( x) .
n n
(2)
information. Compared to Gabor filters, the logarithmic filters
maintain a zero direct component in the even-symmetric filter
The local energy function is defined for signal I (x ) as for arbitrarily large bandwidth filters [22]. This one-
dimensional symmetric/antisymmetric Gabor filter pairs are
E ( x )  F 2 ( x)  H 2 ( x) (3) extended into two dimensions by simply applying Gaussian
as spreading function across the filter perpendicular to its
where F (x) is the original signal I (x) with its direct orientation. By using a bank of these two-dimensional filters,
component removed and H (x) is the Hilbert transform of which is designed to tile the frequency plane uniformly,
features at all orientations can be detected. The energies
F (x) [26]. As shown in Fig. 4, local energy is equal to phase over all orientations are summed and normalized by dividing
congruency scaled by the sum of the Fourier amplitudes by the sum over all orientations and scales of the amplitudes
of the individual wavelet responses at that location x in the
E ( x)  PC ( x) n An ( x ) (4) image. In this way the two-dimensional phase congruency is
Thus, the local energy is directly proportional to the phase calculated as
congruency and energy peaks are corresponding to peaks in (7)
phase congruency, which indicate image feature significance. PC2 ( x ) 
  W ( x) A ( x) ( x)  T 
o n o n, o n,o o

In (1) and (4) the phase congruency is defined as the ratio of   A ( x)  
o n n,o
E (x ) to the overall path length of local Fourier components where o denotes the filter orientation.
(Fig. 4). Thus, the degree of phase congruency is
independent on the overall magnitude of the signal. This B. Feature Histogram
provides invariance to image illumination variations [22]. The Based on the local energy model, a histogram is generated
phase congruency in (1) is sensitive to noise and results in by accumulating the calculated phase congruency (Fig. 5,
© 2011 ACEEE 38
DOI: 01.IJEPE.02.03.92_14


middle) along each filter orientation on different regions of V. EXPERIMENTAL RESULTS
the image (image partition in our experiments is shown in
In this section, we present the experimental results of the
Fig.6). Before histogram conformation, an orientation label
insert inspection, examine the classification rate and also
map for the image is obtained based on local energies. To
evaluate the feature invariance to image transformations.
each pixel a consistent orientation o is assigned, at which it
has largest energy across all scales

L( x )  arg max Wo ( x ) n An, o ( x) n, o ( x)  To  (8)
o

For a single image region, a local histogram is created by
summarizing the phase congruencies in certain orientation.
These local histograms extracted from different regions of
the image are then concatenated together, while keeping the
spatial relationship between image parts. This produces the
histogram for the whole image (Fig. 5, right)
H r , o   x r ( x ) PC2 ( x) ( L( x )  o) (9)
where Figure 6. Image partition and weights of regions for divergence
calculation
1 for L(x)  o
 ( L( x )  o )   (10)
0 for L(x)  o
and r (x ) is Gaussian weighing function centred at region
r . This weight provides soft margins across

Figure 7. A test set of carbide inserts

A. Test Sets
We evaluate the inspection method on real carbide inserts,
Figure 5. Phase congruency-based feature descriptor
which have different geometries from different manufacturers.
histogram bins by small weighted overlap among
An example test set is given in Fig. 7, which contains 14
neighbouring regions to overcome the problems induced by
insert types. For each type there are three insert examples
scale variations [23].
available. We use two examples of each type (namely two
C. Matching test insert sets) to acquire test insert images. For each insert
The classification of chip-former geometry is performed we get with two different scales, under three varied
by matching the histogram of test image independently to illumination conditions and at two random positions 2x3x2=12
the database extracted from template images. For this test images. Corresponding to the insert set and the image
matching the Jeffrey divergence between corresponding scale the test images are grouped in 2x2=4 test image sets.
feature vectors is calculated, which is used as measure of With these insert sets a test for the classification of the
similarity [28] chip-former geometry can be estimated from a sample of
14x2x12=336 test images of real carbide inserts. By using the
R O
hr , o k remaining insert set we acquire one template for each insert
d ( H , K )   r  (hr , o log  kr , o log r , o ) type, which is grabbed under a controlled illumination
r o mr , o mr , o
condition.
(11)
B. Test Results
where hr , o is the corresponding bin in histogram H for In our experiments, local histograms with 4 bins
corresponding to 4 filter orientations on 9 image regions are
region r and orientation o , mr , o is the mean of bins and
extracted, which create a 36-dimensional feature vector. The
given by phase congruency is achieved by convolving the image with
(12) a bank of Gabor wavelets kernels tuned to 4 frequency scales
(hr ,o  k r ,o )
mr ,o  and 6 orientations. In Fig. 8 the test results of the example
2 test set of carbide inserts in Fig. 7 are presented. Our
experiments show, that the descriptor extracted from the phase
and  r  [0,1] is weight of each region and assigns to each
congruency is relatively robust in the production environment
region different influence on the divergence calculation (Fig. and achieves an overall average classification rate of 97.02%.
6).
© 2011 ACEEE 39
DOI: 01.IJEPE.02.03.92_14


It is similarly effective compared to the method developed in image contains more details. These subtleties prepare again
[3], which uses the SIFT local feature descriptor [11], image features, which are not represented in low-resolution
diminishes the false matched feature vectors with a location image. The developed descriptor is sensitive to these
consistency check of feature points and provides a
classification rate of 98.56% for the test on the same image subtleties and therefore the invariance to image scale is
sets. expected only in a closely limited range.

Figure 10. Classification rates for comparison with differently
Figure 8. Classification rates of four test image sets scaled references

CONCLUSIONS
This paper describes a measuring method for an inspection
in carbide insert production. The experiments on real inserts
show that the proposed, on phase congruency-based method
can achieve both accuracy for chip-former geometry
recognition and high robustness to illumination.

REFERENCES
[1] T. Pfeifer, Production Metrology, Oldenbourg, 2002.
[2] R. Hermes, Entwicklung flexibler Bildverarbeitungsketten zur
Klassifikation und Verschleißmessung an Wendeschneidplatten, PhD
thesis, RWTH University, 2007 (in German).
[3] R. Schmitt, Y. Cai, and T. Aach, “A Priori Knowledge-Based
Figure 9. Comparison of phase congruencies under different Recognition and Inspection in Carbide Insert Production”,
illuminations Proceedings IEEE International Conference on Computational
Intelligence for Measurement Systems and Applications (CIMSA
C. Feature Invariance 2010), pp. 339-342, 2010.
In this section the robustness of the feature descriptor with [4] R. O. Duda, P. E. Hart and D.G. Stork, Pattern Classification,
respect to image translation, rotation, scale change and 2nd edition, John Wiley & Sons, 2001.
illumination change is evaluated. In Fig. 9 the phase [5] C. M. Bishop, Pattern Recognition and Machine Learning,
congruencies for insert images acquired under three Springer, 2006.
illumination settings are shown and the invariance to image [6] R. Brunelli, Template Matching Techniques in Computer Vision:
Theory and Practice, Wiley, 2009.
illumination can be indicated exemplarily. The phase
[7] J. Mundy, “Object Recognition in the Geometric Era: a
congruencies show no obvious differences due to the Retrospective”, in Toward category-level object recognition, edited
changed lighting situations and the histograms are also very by J. Ponce, M. Hebert, C. Schmid and A. Zisserman, Springer, pp.
similar. Because the histogram is concatenated after the spatial 3-29, 2006.
relationship of image regions, this phase congruency-based [8] D. Zhang, G. Lu, “Review of shape representation and
feature descriptor is not invariant to translation and rotation. description techniques”, Pattern Recognition, Vol. 37, No. 1, pp. 1-
Our experiments show that the classification rate drops 19, 2004.
[9] H. Murase and S. K. Nayar, “Visual learning and recognition
strongly from a rotation of 5 and from a translation of a of 3-D objects from appearance”, International Journal of Computer
tenth of the image size. In order to analyse the scale Vision, 14, pp. 5-24, 1995.
invariance, test images are compared with differently scaled [10] C. Harris and M. Stephens, “A combined corner and edge
reference data sets. The corresponding classification rates detector”. In Alvey Vision Conference, pp. 147–151, 1988.
are shown in Fig. 10. The blue curves present the results [11] D. G. Lowe, “Distinctive image features from scale-invariant
keypoints”. International Journal of Computer Vision, pp. 91-110,
from the matching with the low-resolution reference
2004.
(horizontal resolution 310 pixels), while red curves illustrate [12] Y. Ke and R. Sukthankar, “PCA-SIFT: A More Distinctive
the rates from the comparison with the high-resolution Representation for Local Image Descriptors,” Proc. Conf. Computer
reference (horizontal resolution 435 pixels). High-resolution Vision and Pattern Recognition, pp. 511-517, 2004.
© 2011 ACEEE 40
DOI: 01.IJEPE.02.03.92_14


[13] K. Mikolajczyk and C. Schmid, “A performance evaluation of [20] A. Hornberg, Handbook of Machine Vision, Wiley-VCH, 2006.
local descriptors”. IEEE Transactions on Pattern Analysis and [21] B. Jähne, Digital Image Processing, Springer, 2002.
Machine Intelligence, 27(10), pp. 1615-1630, 2005. [22] P. D. Kovesi, “Image Features from Phase Congruency”,
[14] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J. Videre: Journal of Computer Vision Research, MIT Press. Vo. 1,
Matas, F. Schaffalitzky, T. Kadir, and L. Van Gool, “A comparison Nr. 3, pp. 1-26, 1999.
of affine region detectors”. International Journal of Computer [23] S. Sarfraz and O. Hellwich, “Head Pose Estimation in Face
Vision, 65(1/2), pp. 43-72, 2006. recognition Across Pose Scenarios”, Proceedings of VISAPP 2008,
[15] G. K. Toledo, E. Kussul, T. Baidyk, “Neural classifier for Int. conference on Computer Vision Theory and Applications, pp.
micro work piece recognition”, Image and Vision Computing, 24, 235-242, 2008.
pp. 827-836, 2006. [24] K. Anusudha1, S. A. Koshie, S. S. Ganesh and K.
[16] A. Shafi, “Machine vision in automotive manufacturing”, Mohanaprasad, “Image Splicing Detection involving Moment-
Sensor Review, Vol. 24, Iss. 4, pp. 337-342, 2204. based Feature Extraction and Classification using Artificial Neural
[17] Y. Wang, W. Fu, H. Zhu and Y. Liang, „A Method for Networks”, ACEEE Int. J. on Signal & Image Processing, Vol. 01,
Recognition of Work-pieces Based on Improved SIFT No. 03, pp. 9-13, 2010.
Characteristics Matching”, Proceedings of the 2009 WRI World [25] M. C. Morrone and R. A. Owens, “Feature Detection from
Congress on Software Engineering, pp. 29-33, 2009. Local Energy”, Pattern Recognition Letters, 6, pp. 303-313, 1987.
[18] M. Peña, I. López, R. Osorio, “Invariant Object Recognition [26] S. Venkatesh and R. A. Owens, “An Energy Feature Detection
Robot Vision System for Assembly”, Proceedings of the Electronics, Scheme”, International Conference on Image Processing, pp. 553-
Robotics and Automotive Mechanics Conference, pp. 30-06, 2006. 557, 1989.
[19] E. Kussul, T. Baidyk, D. Wunsch, O. Makeyev and A. Martín, [27] K. Gröchenig, Foundations of Time-frequency Analysis,
“Image Recognition Systems Based on Random Local Descriptors”, Birkhäuser, 2001.
Proceedings of Joint Conference on Neural Networks, pp. 2415- [28] Y. Rubner, C. Tomasi and L. J. Guibas, “The earth Mover’s
2420, 2006. Distance as a Metric for Image Retrieval”, International Journal of
Computer Vision, 40(2), pp. 99–121, 2000.

© 2011 ACEEE 41
DOI: 01.IJEPE.02.03.92_14

Illumination-robust Recognition and Inspection in Carbide Insert Production

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to Illumination-robust Recognition and Inspection in Carbide Insert Production (20)

More from IDES Editor (20)

Recently uploaded (20)

Illumination-robust Recognition and Inspection in Carbide Insert Production