A Novel Defect Inspection Method for the TFT-LCD Image Based on ...

at once. The tri-modal thresholding function fT ( x,
y)
is
given by
f T
255
if ( f ( x,
y)
m k
)
( x,
y)
128
else if ( f ( x,
y)
m k
)
0
else
(10)
Where, f ( x,
y)
is <strong>the</strong> gray level at position at ( x , y)
, m ,
are <strong>the</strong> mean and standard deviation respectively and k is
controlled constant to access <strong>the</strong> amounts of abnormal gray
level. After tri-modal thresholding <strong>the</strong> Bright Muras have
value of 255, <strong>the</strong> Dark Muras have 128 and background
region have value of 0.
3.4 <strong>Defect</strong> Analysis
The segmented result can have false defects such noise,
neglectful small region, etc. So reliable defect confirmation,
defect analysis process is needed. In this paper, HVS based
false region elimination method is proposed. The noticeable
Luminance difference is given by <strong>the</strong> Weber Constant C Weber
in <strong>the</strong> Weber region as <strong>the</strong> following Equation (11) [11].
C W eber
L
(11)
L
Where, L, L
are <strong>the</strong> background region luminance and <strong>the</strong>
luminance differences between <strong>the</strong> background region and
defect region. We use C value as 0.02 in this paper.
Weber
4 Experimental Results
The proposed method is tested by using <strong>the</strong> 160 real
Muras and 40 generated ones. The acquired <strong>TFT</strong>-<strong>LCD</strong>
images have 400 m spatial resolution <strong>for</strong> each pixel and an
8-bit brightness resolution. To evaluate <strong>the</strong> per<strong>for</strong>mance of
proposed algorithm, one dimensional line profile, <strong>the</strong>
enhanced result image and final segmented Muras along
with detection accuracy table are shown in this paper.
The proposed defect detection results are shown in Fig. 7.
The test image Fig. 7(a) has artificially generated Muras
with various strengths and sizes. In <strong>the</strong> enhanced result
image, Fig. 7(b), <strong>the</strong> defect is more visible than original
image and that are evident in Fig.7 (e),(h),(k) also. The
original image of Fig. 7(g) has much background variation,
so <strong>the</strong> defects are rarely seen in <strong>the</strong> original image. But <strong>the</strong>
enhanced image has relatively flattening background and <strong>the</strong>
defects are more salient. In <strong>the</strong> test image Fig. 7(j), novel
enhancement result is given and that helps final
segmentation result.
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
(j) (k) (l)
Fig. 7 The results of proposed methods. (a),(d),(g),(j) are test images,
(b),(e),(h),(k) are enhanced results and (c),(f),(i),(l) are final segmented
results of original image (a),(d),(g),(j) respectively.
The microscopic enhancement results by <strong>the</strong> line profiles are
shown in Fig. 8. Fig. 8 is <strong>the</strong> enhancement results’ profiles
of Fig. 7. By that, <strong>the</strong> effectiveness of our algorithm is seen
more clearly. In <strong>the</strong> filtered line profile results, our proposed
enhancement method can flatten <strong>the</strong> background signal
fluctuation and emphasize <strong>the</strong> defect region.
The final per<strong>for</strong>mance results are summarized in Table 1.
false-positive and false-negative measure are used to
evaluate <strong>the</strong> accuracy of defect detection result. There are
some problems to <strong>the</strong> elimination of defects existing at <strong>the</strong>
boundary region and very weak-small-size defects. Except
such cases, <strong>the</strong> false-positive rate and false-negative rate are
quite acceptable.
Table 1. Experimental results <strong>for</strong> <strong>the</strong> detection of <strong>the</strong>
Muras.
Number of
<strong>Defect</strong>s
False-positive Falsenegative
Artificially
Generated Mura
160 0.04 0.03
Real Mura 40 0.03 0.02