Instrumentation & Measurement Magazine 24-4 - 62

Fig. 3. Variation in the SSE according to the polynomial fitting order. (a) The SSE for different orders; (b) A partially
enlarged picture of the emphasized square in (a).
It is:
Gx( , ) (, )
y G x y W
(5)
where G0(x,y) represents the gradient magnitude of the current
pixel; and W is the threshold of the gradient magnitude. When
the absolute difference of the gradient amplitude exceeds a defined
gradient magnitude threshold, the point is fixed as the
edge point and the growth stop; otherwise, the point rejoins
the growing region as a new seed. By adding the growing stop
condition, the growing condition can be appropriately relaxed,
and the threshold T can be increased to prevent under-segmentation
caused by grayscale changes.
Processing Results
Fig. 4. Experimental results and fitted results for ice defects.
determining the location and edge of any defects. An adaptive
Canny edge detection algorithm based on morphological filtering
and an Otsu algorithm were used to locate the defects.
The morphological filtering was used to remove noise and retain
good edge information. This also improved the efficiency
of the algorithm. After calculating the amplitude and direction
of the gradient by adopting a first-order partial derivative finite
difference method and non-maximal suppression of the
gradient's amplitude, the Otsu algorithm was employed to
compute the upper and lower thresholds [11]. This enabled the
final edge to be obtained, which served as a basis for selecting
seed points (points in a target object) to further identify the defective
area. There can be too many centroids on the edge of
each segment if centroids are adopted as the seed points for
edge detection. A clustering method was therefore required to
cluster the centroids and remove the seed points with larger
differences. A dual thresholding criterion was also applied to
prevent under- or over-segmentation during the region growing.
The first criterion for an edge point to serve as a growing
condition is:
f (, )xy Q T
62
(4)
Four samples of honeycomb sandwich skin with ice defects,
labelled A, B, C, and D, were subjected to detection and processing.
The infrared thermal images processed using the combined
methods described above are shown in the bottom row in Fig. 5.
The original experimental images are in the top row.
The defective regions and the effect of the improved region
growing method on the defective region extraction can be
quantitatively evaluated using the following:
SS
N
dt
t




N w
where Sd is the defective area; St
Nw is the number of white pixel points; and Nt
ber of pixel points.
(6)
is the total area of the image;
is the total numTable
2-Evaluated area of the ice defects
Sample
A
B
C
D
Se/cm2
38.58
41.35
11.96
8.18
IEEE Instrumentation & Measurement Magazine
Sd/cm2
37.61
41.82
11.66
7.68
Relative
Error

Instrumentation & Measurement Magazine 24-4

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-4