bottom corners of a specimen). Matlab's 'improfile' command was then used to draw a
line between these two points and the number of pixels this line passes through was
counted. The known distance divided by the number of pixels can then be used as the
length ratio for each pixel (measured in mm/pixel).
Figure 6-8 illustrates the boundary trace method applied to Defect A75 and Defect
IB on Specimen A-1. The first step in the analysis procedure is to identify the image that
was collected at tmax (time of maximum defect signal strength). The color scale of the
image is then adjusted such that the entire scale is distributed across the range of
temperature values encountered in the box used to define the defect. Next, Matlab's
"roipoly" command is invoked and the user traces out the boundary of the defect in the
thermal image. The results of this analysis for Defect A75 (19 mm diameter air-filled
defect) are provided in Figure 6-8A. The total number of pixels bounded by the trace
was 377. After applying the length factor for this image (1.1 mm/pixel or 1.2
mm2/pixel), the area of the defect was estimated to be 4.4 cm2. The true area for this
defect was 2.8 cm2. The same procedure applied to Defect IB on Specimen A-i resulted
in an estimated area of 10.6 cm2. The true area for this defect was 1.2 cm2.
Further experimentation with other defects of known size indicated that the
boundary trace method consistently overestimates the size of the defect. It is conceivable
that this bias error could be quantified and then considered in future computations. The
fact remains, however, that selecting the boundary of the defect will always require some
degree of human judgment. Maldague (2001) outlines a procedure for approximating the
size of a defect by computing the magnitude of the maximum temperature gradient at
each pixel in a thermal image. The underlying principle for this procedure is that the