Abstarct: Recognizing sign languages are widely used for communicating deaf people and people who hear hard with other people or with computer. Nowadays, recognizing by using images processing tools have been interested by researchers, because of its importance. In this research we want to recognize hints and hand signs by using 2-dimensional image processing without using any special equipment. One of the best methods for identifying hints and signs is to comparing hand motion path and modeling hand shape and how of locating its fingers. We introduced and investigated maximum edge direction (MED) in our research and indicated that MED is a geometric feature that is related to hand structure. MED feature, statistically, indicate maximum direction of image edges and indicate direction of locating fingers in each hint, accurately. For extracting this feature, first we take a binary image of hand image edges and then appropriate proposed structural element, which works only on edge pixel, are used. Also for a better modeling of hand shape, LMED feature are used. Then target images are partitioned to smaller region with definite size and MED of each region are calculated. Using this feature strongly decrease the amount of needed databaxse and increase recognizing rate. Proposed algorithm for recognizing sign language contains 2 main steps: 1-initial classification 2- final classification. In first step, features like: hand motion path and curves of hand image's area and hand sinew angle are extracted. Then undesirable classes are removed by using DTW algorithm and nearest neighbor classifier. Finally 3 classes remain that right class is certainly between these classes. Consequently speed and accuracy of sign language recognition are improved. In second step ,first by using DTW algorithm, that curve of hand motion path is it's function, same frxames are selected and then distance between input sign and 3 classes ,in first step, are calculated by using two features :LMED and hand motion path. In this project 3 methods are proposed for sign language classification: 1- Minimum distance method by using LMED feature. 2- Minimum geometrical average method 3- Minimum Euclidean distance method adaptively. Proposed classifiers are tested on a collation of ASL sign that contain 47 different words with 137 samples. In this experiment we reached to accuracy rate of %97.7 that are about %98.9 in best situation for hinting with one hand and about %95.6 for hinting with 2 hands. Results indicate that proposed method has a high efficiency