Abstract:The problem of handwritten recognition is considered to be very important problem because of its numerous applications and theoretical values in the domain of pattern recognition. This thesis presents a design to an off-line handwritten recognition system. It is designed to recognize handwritten styles of Latin script. The recognition responsibility of the proposed system is for 62 character classes [uppercases (A-Z), lowercase (a-z) and digits (0-9)]. The suggested system includes the essential stages needed for most of the pattern recognition systems. These stages are the preprocessing stage, the feature extraction stage, the pattern matching and classification stage. The design work is for approaches of the steps of the preprocessing stage like the information removal step of the adaptive thresholding resultant characters, the contour tracking algorithm, and the interpolation step of both the contour slope function and the complex plane method to uniform number of points representing characters. The design also includes Fourier and wavelet bases feature extractors. The discrete Fourier and wavelet transforms are used to extract the features vectors of the handwritten character patterns that are required for the classification stage. As a wavelet function, the Haar wavelet has been used. The classification stage is implemented using two methods: the minimum distance classifier depending on Euclidean Distance which has a high performance and the cross-correlation as a second classification method.The design task is in two phases. The first one is the training phase which includes the handwritten collection of the training samples, some necessary preprocessing steps and the feature extraction for all classes. At the end of this phase the feature vectors are ready to the second phase that is the recognition phase. The system was implemented using MTLAB program version 7, database of 289 character samples was established, 248 samples from the data base were used for training and 41 samples were used for testing the system. The efficiency of the system was tested and results show that the usage of wavelet transform feature extraction and minimum distance classifier achieves the highest performance in which it scores an reliability of 90.24%.