Abstarct: OCR plays an important role in different areas such as license plate recognition, security (passport authentication), barcode and bank systems (check). There are many methods and systems for recognizing characters. The purpose of this study proposed a method to recognize letters and digits in a language free limitation in font type and size. The most existing recognition systems have many training samples and as a result, their computation complexity is high; because they use a set of training samples per font. Also, if these systems have been extended to identify more fonts, the recognition error rate is increased. So, it needs a system with small training samples which can be resistant to different the fonts and Fault diagnosis not much has changed by adding new fonts. In this thesis, we recognize typed digits and letters in English. Typed characters are written in different fonts and sizes that make similar characters being different with each other; therefore, a system can recognize similar characters with extracting appropriate features form the characters. In this study, we could identify English digits in 30 different font and 11 different sizes with 99.47% accuracy and English letters in 24 different font and 11 different sizes with 97.73% accuracy using a simple feature vector- the sum of values of pixels in each row and column- and a SOM neural network without complex architecture. Self-Organizing Map (SOM) usually uses Euclidean distance to compare the samples of a class. But in this thesis, instead of using Euclidean distance, we use similarity measure to compare the feature vector of input digit and different neurons’ weight. One or two samples are used to train the network and it doesn’t need to have a set of samples in different fonts.