Abstarct: Abstract Automatic recognition of text is a subset of image processing, which widely used in digital library, reading e-mail, reading check bank, reading forms and questionnaires . manuscxript text recognition can be performed in two general states: analytical method and holistic method . In systems baxsed on analytical method, words are divided into their components, sub words and letters which are used to identify word; but in holistic method features are extracted directly from the whole word image and by comparing them wiht the data dictionary (lexicon), the word which is similar to the input is considered as the output. In Preprocessing stage for estimating the location of baxseline, the changes are performed in horizontal histogram method. This process is needed for two reasons. First to isolate the main body of the word and second for normalizing the main components of word. This lead to decompose a word into two parts main body image and diacritics & dots image. In order to describe the image of word After normalizing each component ,we propose to divide the image of main body using adaptive blocking. For second image part (diacritics and dots) we propose using constant blocking. Each block in each of these two images are described using image gradient in 32 directions. The result of experiments shows that the recognition rate of proposed method is improved using k nearest neighbor and support vector machine classifiers. However, because of preprocessing the proposed method require more time than previous methods. The clustering of samples in order to lexicon reduction will improve the results so Horizontal histogram and profile feature with using isodata algorithem are used to cluster samples .