Abstract : The global development of security system around the world made the field of recognition most needed, and led the idea of recognition grew up in researchers work; many specifications were found in mankind to differentiate between one and another. However, some of them were old like finger print and other types were modern as iris print, face, sound, palm print and DNA print. This thesis proposed a face recognition system. It is based mainly on classifier unit, which had been chosen as Radial Basis Function Neural Network (RBFN). This type of image is consist of a huge number of pixels (=width _ height), such large number of pixels make process of classification is so difficult. Pseudo Zernike Moment PZM was used to normalize amount of date that represent face image by extracting the overall description of facial expression in image; before implementing PZM, skin map algorithm employed to normalize the size of image. This was illustrated by neglecting all regions of image except that contains skin. Different sizes of feature vectors were extracted and a comparison between them was used to find optimum size which give best representation by smallest time, 36 features were extracted for each image. In order to solve problem of misclassification problem, hidden layer of RBFN was designed to have 9 neurons, learning rate z and confidence coefficient b were assumed to be 1.414, 0.707 respectively. Proposed system considered as a security system for a company that has 30 employees working in 4 departments, system will read image of each visitor and decide what his department. For each person 3 images with different pose were taken, for totally 90 images ,each one was normalized, its feature was extracted, and then the feature vector of each one was summed to construct an overall training vector T V . Total time t consumed for extracting T V was 340.290 s, and the operation of learning RBFN completed in 15.749 s. A modification was applied to that vector to reduce its size by neglecting zeros that it contains; this process reduces its size to be 15 elements instead of 36; that_s made operation of learning RBFN consumes time of 10.969 s. RBFN was trained by this vector then it was tested by 10 different images to give recognition rate of 90.