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

Title:	Multi oriented text detection and recognition via adaptive binarization and fuzzy clustering
Authors:	Saad Mohmmad Saad Ismail (P63453)
Supervisor:	Siti Norul Huda Sheikh Abdullah, Assoc. Prof. Dr.
Keywords:	Binarization Fuzzy clustering Optical character recognition
Issue Date:	28-Aug-2016
Description:	Document image analysis and recognition or Optical Character Recognition (OCR) refers to algorithms and techniques that are applied for obtaining a computer-readable description from a set of text pixel data. The handwritten ancient and historical script documents have the most complex and unsolved issues like broken characters and poor readability suffered from ink bleed, non-smoothing and extreme low contrast problems. Unfortunately, contemporary simple enhancement techniques are still less effective in restoring most of the text body due to high variation between its text border and surrounded background pixels. Besides, the available compound binarization methods are less applicable for segmenting text in degraded document images caused by illumination, low contrast, thin pen strokes, and non-uniform problems. Those present methods are manually affected by some critical factors like window size and parameters values. Moreover, major cutting-edge challenges in which less attention given to OCR developers are various font type, size, color, background variation like complex noise in natural scene, and text perspective for example horizontal, merely horizontal and multi-orient texts. Inspire from aforementioned motivation, four objectives are highlighted, first is to propose an adaptive enhancement method for historical handwritten manuscript images, second is to propose a compound binarization method for degraded document images, third is to propose a text extraction technique via compound binarization and fuzzy clustering methods and finally is to develop a new OCR framework for natural text scene images. The methodology divides into four phases. At first, the proposed handwritten historical manuscript image enhancement is constructed based on smoothing and contrast filters, hybrid thresholding and noise reduction sub-processes. Secondly, an introduction of proposed compound adaptive binarisation method is rationalized based on the dynamic parameter, window size and hybrid thresholding. Thirdly, the proposed text detection and extraction algorithm is performed based on exploited NTSC enhancement, hybrid segmentation approach comprising Fuzzy C-mean clustering and adaptive compound binarization methods, connected component labelling and stroke width transform subsequently. At final stage, a multi-oriented OCR framework is developed for natural scene image. The experimental results show that the proposed historical manuscript enhancement method produces the text more readable view with RAE and F-measure achievement approximately 0.044 and 76% for Jawi manuscript and Bickley Dairy Dataset respectively. Next, the proposed adaptive compound binarisation method was tested on DIBCO benchmark datasets. The experiment shows that the proposed adaptive compound binarization method (91%) outperforms state of the art methods namely Bataineh, Niblack, Sauvola, NICK and Lazzara_MS_k binarisation methods with average F-mean of 87%, 39%, 78%, 75% and 76% subsequently. Then, the proposed text detection and extraction gained average precision and recall rates about 82% and 89% subsequently for ICDAR and KAIST scene image datasets. Finally, the proposed natural scene OCR framework achieved recognition rates up to 82.23% and 68.06% for ICDAR 2013 and KAIST datasets successively.,Certification of Master's/Doctoral Thesis" is not available
Pages:	204
Call Number:	TA1640.I844 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.