This paper presents an automatic off-line handwritten cursive Arabic recognition system. The system is based on an artificial neural network classifier. The preprocessing step includes binarization, noise reduction, and thinning. A new thinning algorithm is developed that produces a skeleton of the characters without gaps or extra branches. The proposed word segmentation approach is based on following the base line of the thinned word or sub-word, the base line is calculated by analysis of horizontal density histogram. In the recognition stage, four different sets of characters have been independently considered which are: isolated, beginning, middle, and end. A neural network is used for each set. The neural network uses the principle component analysis PCA as a tool for feature extraction. Where it compresses the character to a certain number of features (feature dimension). The classification is done by MLP neural network trained with back-propagation. The system has been tested and has shown a high accuracy.