Imaging and image processing have been used in variety of applications, such as medical,
astronomy, forensic, and industry. Numerous techniques have been proposed to solve
speci c problems faced in particular applications which are comprised of a series of processes
such as, image enhancement, ltering, segmentation, representation, and recognition.
However, there is no a universal algorithm which can be applied to variant image modalities
with corresponding applications.
With the aim of learning image processing tasks, as supervised learning techniques, we
can develop e ective algorithms which are image and task oriented. Learning image processing
comprises two main phases, namely: training and testing phases. During training
phase, the algorithm has the capability of discovering and adjusting an optimum transformation
function or optimal mathematical morphology chain by utilizing a user-prepared
ground-truth (gold) sample. Later on, in testing phase, the obtained transform function
of morphological chain is applied to untrained test images.
The current thesis has three main parts as follows. In the rst part, genetic programming
(GP) is employed to obtain an optimum transformation function. The GP utilizes
one user-prepared gold sample to learn from. The magni cent feature of this method is
that it does not require a prior knowledge or large training set to learn from. The performance
of the proposed approach has been examined on 147 X-ray lung images. The
results for image thresholding (i.e., Otsu's method) after applying optimal transformation
are promising.
In the second part, an optimum mathematical morphology (MM) chain is obtained by
applying GP to localize the object of interest in a binary image. Morphology operations
use 27 regular structuring elements along with commonly used morphological operations
(i.e., erosion, dilation, opening, and closing) to build an optimal MM chain. The obtained
chains are tested against challenging test cases, such as, object translation, scaling, and
rotation.
In the third part, a hybrid genetic programming - di erential evolutionary (GP-DE)
algorithm is proposed to optimize not only the morphology chain but also the utilized
structuring elements. GP as an outer layer optimizer is responsible to optimize the morphology
chain while the di erential evolutionary (DE) as an inner layer optimizer optimizes
the structure elements. Similarly in the testing phase, the obtained morphology chain is
applied on test images. In term of utilized test images, the two test cases have been employed
: synthesis and music note images. The results indicate that the proposed method
is able to locate the object of interest. For the music note images, the proposed approach
is able to extract the head notes, sta s, and vertical lines correctly. The training phase is
iii
time consuming, but it is acceptable; because one time training is required to obtain an
optimal chain for a speci c image processing task.
| Identifer | oai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:OOSHDU.10155/424 |
| Date | 01 March 2014 |
| Creators | Shahbazpanahi, Shaho |
| Contributors | Rahnamayan, Shahryar |
| Source Sets | Library and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada |
| Language | English |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0015 seconds