Thesis (M.Sc.) - Cairo University - Faculty Of Science - Department Of Mathematics

In the proposed system , the scanned image is processed through five consequent phases ; preprocessing , segmentation , feature extraction , classification , and binding phasesIn the preprocessing phase , median filter is used to eliminate noiseSegmentation phase utilizes the upper and lower contours of each Part of Arabic word (PAW) to define segmentation points using a set of segmentation rulesThis segmentation algorithm is capable of segmenting overlapped , overhanged , and touched charactersThe output of segmentation process is a list of separated primitives in conjunction with their secondariesFeature vector (FV) of each primitive is obtained using 8 directions chain code approachEach primitive is projected on a circular grid centered at the center of gravity of the primitive to achieve translation invariant FVThe contours chain codes of primitives parts lying in different gird segments are concatenated together to form the primitive FVNearest neighbor and neural network (NN) classifiers are used to classify primitivesNearest neighbor classifier achieves higher primitives recognition ratio than the NNGeometrical based classifier is used to classify secondaries into four different classes (dot , two Dots , three dots , and Hamza) In the binding phase , characters and words are completely recognized by linking different primitives belonging to the same character together and associating the assembled character with its corresponding secondariesBinding phase improves

Issued also as CD