1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Improvement of gradient vector flow using particle swarm optimization for broken character restoration and classification
Authors:	Qusay Omran Mosa Al-Mramthee (P64768)
Supervisor:	Mohammad Faidzul Nasrudin, Assoc. Prof. Dr.
Keywords:	Vector Particle Character Classification Neural networks (Computer science)
Issue Date:	Feb-2016
Description:	Degraded ancient documents contain unrecognized symbols and broken characters. This becomes a challenge to optical character recognition (OCR) since they tend to be identified as noises. Current methods of restoring broken characters are centered on the renowned Active Contour or Snake algorithm. Gradient Vector Flow (GVF) Snake, an extension of the Snake, is proved to be an effective method because it able to identify concave boundaries. However, it fails in exploring deep or narrow concave boundaries and consumes high computation. The thesis introduces three new methods and algorithms to improve the GVF Snake. Firstly is to propose the use of T-triangle Steps and Balloon Force algorithms in the GVF Snake. T-triangle Steps algorithm helps to detect deep concave regions and then Balloon Force algorithm inflates inside the region until points of Snake (snaxels) positions is detected. Secondly is to extend the convergence search space by proposing a Genetic-GVF Snake algorithm. Genetic algorithm is an optimizer that guides snaxels from GVF Snake to handle very complex and long deep concave boundaries. Thirdly is to propose the use of Particle Swarm Optimization (PSO) with GVF Snake to speed up the computations and decrease the iterations of the algorithm. Then the classification operation is done using multi layer feed forword neural networks and tained by error back propagation to ensure an efficient solution of handwritten pattern recogniyion. The PSO uses the Tournament Selection algorithm to accelerate its particle selection time and avoid trapping in local minimum. All broken characters are produced from the ISO and MNIST datasets which has a total of 60,026 character images. Results from the proposed methods are compared with results from the traditional algorithms of respective proposed methods using the Hausdorff Distance (HD) as the image matching performance matrix. By using the Triangle steps with Balloon force, most of broken characters are able to be restored with a rate of 96% with an improvement of HD average from 23.214 to 0.586. Besides that, the applied Genetic-GVF Snake led to an enhancement of the restoration rate to 100% for broken characters and enhancement of the average of restoration value (HD) to 0.3583. As compared with the Genetic-GVF Snake, the improved PSO-GVF Snake is proved to decreases the average number of iterations needed to restore the broken characters by 2.5363 times faster than iteration needed using the Genetic-GVF Snake. The results show that the traditional Snake need some improvements to make it practical for broken character restoration application. The Genetic-GVF Snake is able to solve the problem of long deep concavity for some object boundaries and restored all images of broken characters. The final PSO-GVF Snake algorithm is able to restore broken character quicker than the other methods in such a way that the broken characters can be recognized with a high recognition rate. The training using neural networks provide a better result of recognition by get the classification rate of %95 of MNIST and ISO datset.,Certification of Master's/Doctoral Thesis" is not available
Pages:	197
Call Number:	QA76.87.A447 2016 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.