Instrumentation & Measurement Magazine 25-2 - 10

information and other information, classify the raw data into
several categories and then highlight the category of defect
information. The most common statistical methods in ECPT
sequences processing are binary segmentation, independent
component analysis (ICA) and so on.
Binary segmentation is based on spatial temperature distribution
differences between defect area and non-defect areas.
Through an appropriate threshold, each pixel of the image
would be set to 0 (if the pixel's value is lower than the threshold)
or 1 (if the pixel's value is higher than the threshold). The
threshold is generally set from the perspective of statistics. For
example, Otsu algorithm [10], a widely used global threshold
algorithm, sets the threshold by maximizing the between-class
variance (BCV) between the foreground and background after
segmentation. Niblack algorithm [11] and Sauvola algorithm
[12], two famous local threshold algorithms, obtain binarization
threshold by calculating the mean value and standard
deviation of a pixel and its neighborhood pixels. K-means
clustering algorithm [13], an iterative algorithm for images
segmentation, keeps updating pixels to k different clusters
via the minimum Euclidean distance, until the optimal lowest
sum of squares of errors. In addition, there are also recent
studies combining machine learning algorithm and first-order
statistical characteristics. For example, in [14], the maximum
inflection point of the bins of frequency histogram is used as
the global threshold to segment the defect and background,
and the number of the bins is adapted to adjust by genetic algorithm,
which achieved better results than traditional methods
such as OTSU and K-Means (Fig. 6 [14]).
Binarization segmentation algorithm runs fast and can
quickly distinguish high and low temperature areas. However,
it also has drawbacks when applied in ECPT. The binarization
method processes single frame images rather than image sequences
and only considers the absolute value of temperature
without considering the relationship between temperature
and time. In fact, in addition to the defect area, some nondefect
areas, such as coil and specimen edge, also have high
temperature. Binarization segmentation algorithm makes it
hard to distinguish the differences between such kinds of nondefect
areas and defects.
ICA is a blind source separation algorithm which classifies
raw data automatically without prior information. ICA is
based on the principle of independence in statistics, that is, it
assumes that a signal (marked as
XR 

KN
) can be decomposed
into a linear combination of several statistically independent
components:
X = AS
(10)
where X is the raw data, N is the number of independent components,
S is a matrix whose components (column vectors) are
statistically independent, and A is the mixing matrix. Take an
ECPT image sequence as an example, which is composed of
linear combination of defects, coils, edge noise and other noise
with different features. ICA aims to extract these independent
components from the ECPT image sequences. FastICA [15] is
a most common ICA algorithm for solving (10), which aims to
find a direction in which the projection of data has the maximum
non-Gaussian property (the non-Gaussian property can
be measured by kurtosis or negentropy). It is worth noting that
before FastICA decomposition, the data should be whitened
to ensure that each component of the data is unrelated and the
variance is 1. Specifically, the data is normalized first, and then
processed by principal component analysis (PCA), and then
each component in the PCA
result is normalized.
A typical example of
Fig. 6. (a) Test specimen; (b) Raw ECPT image; (c) OTSU segmentation; (d) K-Means segmentation; (e) FSOP-GA
segmentation. (From [14], ©2018, IEEE).
10
IEEE Instrumentation & Measurement Magazine
ICA's application in ECPT
is to utilize the temporal
and spatial variation,
preset different principal
components and separate
the principal component
corresponding to defects.
For example, [16] deeply
analyzed temperature patterns
and: divided the
ECPT images into four
parts: the area around
crack tips, the side of the
cracks, the area underneath
the coil, and defect
free area; highlighted small
natural thermal fatigue defects;
and suppressed the
coil noise effectively (IC4
of Fig. 7b). In addition, ICA
is suitable for combining
April 2022

Instrumentation & Measurement Magazine 25-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-2