Astatistical performance in dicator in some image processing problems / Chang Yun Fah
The ability to compare or relate two digital images may be useful in developing performance evaluation algorithms. This thesis investigates the use of a particular correlation measure, 2 p R developed from the multidimensional unreplicated linear functional relationship (MULFR) model with single...
Saved in:
| Summary: | The ability to compare or relate two digital images may be useful in developing
performance evaluation algorithms. This thesis investigates the use of a particular
correlation measure, 2
p R developed from the multidimensional unreplicated linear
functional relationship (MULFR) model with single slope, as a measure or indicator of
performance. This MULFR model is an extended version of the ULFR model
introduced by Adcock in 1877. A literature survey was carried out showing that 2
p R has
not been used before. The coefficient 2
p R was investigated in its ability to handle the
issues of non-perfect reference image, multiple image attributes and combining image
local-global information simultaneously. This survey is followed with the maximum
likelihood estimation of parameters and a brief discussion of some theoretical properties
of 2
p R . To investigate robust properties of 2
p R , an extensive simulation exercise was
then carried out. Promising results, thus far, motivate the use of 2
p R in two image
analysis problems; firstly a character recognition problem and secondly a particular data
compression problem. In a handwritten Chinese character recognition problem, the 2
p R
achieved the highest recognition rates even the pre-processing stage is removed from
the recognition system. A substantial reduction of processing time, approximately
40.36% to 75.31%, was achieved using 2
p R . In JPEG compression problem, 2
p R is used
as a measure of image quality which in turn indicates the performance of the
compression method. It is shown that 2
p R performs well and satisfies the monotonicity,
accuracy and consistency properties when perfect reference image was used. 2
p R was
also shown to perform better than some frequently used similarity measures when
imperfect reference image was used. |
|---|