and small. The image quality is deranged by mud,
powder, grease or water and the lighting conditions
change frequently. Therefore, the choice of features
for the detection task is important. Recent approaches
for defect or anomaly detection focus on fault de-
tection in material-surfaces. In (Platzer et al., 2008)
we introduced a one-class classification approach for
anomaly detection in wire ropes using linear predic-
tion (LP) coefficients as features and a Gaussian mix-
ture for model learning. The former work is ex-
tended in this paper by two main aspects: the per-
formance of LP features is compared to that of other
well-established features from the field of textural de-
fect detection, and the robustness to outliers in the
training set as well as the generalization ability of
the presented approach are carefully evaluated. The
last point is of particular interest for the practical rel-
evance of the method.
Features based on local binary patterns (LBP)
were first introduced by Ojala (Ojala et al., 1996)
for texture classification. Recently they were used
for defect detection in fabrics (Tajeripour et al.,
2008) and for real-time surface inspection (M¨aenp¨a¨a
et al., 2003). Textural features, extracted from co-
occurrence matrices, were proposed by Harlick in
the early 70’s (Harlick et al., 1973) and are fre-
quently used for texture description (Chen et al.,
1998). (Iivarinen, 2000) compares two histogram-
based methods for surface defect detection using LBP
and co-occurrence matrices. (Rautkorpi and Iivari-
nen, 2005) used shaped-based co-occurrence matrices
for the classification of metal surface defects. (Vartia-
ninen et al., 2008) focus on the detection of irregular-
ities in regular, periodic patterns. They separate the
image data in a regular and an irregular part. Based
on the resulting irregularities, we compute local his-
tograms, which serve as features. Another important
category of features for texture analysis and textural
defect detection are wavelet-based features. (Kumar
and Pang, 2002) for example use Gabor features for
the detection of defects in textured material. How-
ever, the computation of these features requires large
filter banks and high computational costs. Due to
the huge size of rope data sets (20-30 GB) the time-
consuming computation of Gabor features seems to
be not the best choice. In (Varma and Zisserman,
2003) the authors state, that similar results to that
obtained by the usage of wavelet features can be re-
solved with help of joint neighborhood distributions
and less computational effort.
The one-class classification strategy proposed in
(Platzer et al., 2008) was chosen due to a lack of de-
fective training samples for supervised classification.
In contrast, it is no problem to design a huge sample
set of faultless training samples. With this faultfree
training set a model of the intact rope structure can
be learned and in the detection step outliers with re-
gard to this model are classified as defect. However,
the only available ground truth information about this
training data is the labeling of the human expert. In
the following,there remains a small uncertainty of un-
derdiagnosed defects in the training set. For this rea-
son the robustness of the proposed method to outliers
in the training set is evaluated. Results obtained by
learning from a faultless training set are compared to
those, obtained by learning from a training set with
intentionally added, faulty samples. The generaliza-
tion ability of a learned model is a further important
point, especially for the practical relevance of the pre-
sented method. There exist only a limited number of
different construction types for wire ropes. The dif-
ferences between them are mainly a different num-
ber of wires and strands, different thickness of single
wires, the length of twist and the diameter. If just one
model for every possible rope type would have to be
learned in advance, this would save a lot of compu-
tational effort. However, the rope data from different
ropes differs significantly due to the changing acquisi-
tion conditions and a different mounting of the ropes.
Nevertheless, it is desirable to have just one model
for every construction type and to overcome the chal-
lenges of a changing acquisition environment. There-
fore, the generalization ability of the learned models
is evaluated by learning and testing on different rope
data from nearly identical constructed ropes.
The paper is structured as follows: in section 2 we
briefly review the feature extraction using linear pre-
diction and give a description of the used textural fea-
tures and their extraction. The one-class classification
of wire rope data is shortly summarized in section 3.
Experiments, revealingthe usability and robustness of
our approach, have been performed on real-life rope
data and results are presented in section 4. A conclu-
sion and a discussion about future work is given in
section 5.
2 FEATURE EXTRACTION
In this section the different features are briefly re-
viewed. Their extraction from the underlying rope
data is described, as it differs for the LP features in
contrast to the remaining features. In the following, a
short motivation for the different features used in this
context, is given.
Local binary patterns (LBP) code the local
graylevel-structure of a pixel neighborhood. His-
tograms based on the resulting codes lead to a local
VISAPP 2009 - International Conference on Computer Vision Theory and Applications
172