Abstarct: Nowadays qhe use of biological characteristics to authenticate individuals is increasing. The most biological characteristics which used in this field are fingerprint, iris, retina, face, voice and signature. In the administrative System, especially in the field of economic and commercial sectors, such as banks, corporations, organizations are seeking ways to identify people. One of these tools which are used a lot cause of its simplicity and short cost is signature, But individual's signatures is at risk of forge by other people. Therefore an efficient method to recognize genuine signature from forgeries would be useful. According to the dependency to the kind of receiving the data, signature verification can be divided into two groups, online and offline. In online methods, Signatures information and the information of the dynamic characteristics of signatures such as pressure and speed and ... can be achieved by the use of a tablet and an electric pen which is connected to a computer. While in offline approach, information where collected when signing opperation is done, Then at the other time written signature is converted to digital image by a scanner. In this case, qhe extracted features are called static. Offline systems are most practical in comparison with online systems and are used in a simpler way in all over the world. However because of the laBk of qeO dynamic information, they are supposed to be more BEmplicated against online systems. In this thesis our work is baxsed on the offline systems. At first in order to preprocessing we cropoed the signatures, and then by a special segmentation on signature image, new gride-baxsed local features are exctracted.An continue optimal subset of these features are selected by using Imperialist Competitive Algorithm. In classification level, an artificlui Neural Network is used for each signer which we consider 22 signatures for training phase. In order to evaluation phase, the 22 remained signatures of each individual are injected to it's corresponding neural netwo k which eun rOOn iOu nOa in the previous step. Decision about refusing or acceptqlnd of the test signature is made. Finally the results are compared with similar projects. In this thesis we used GPDS311GARY databaxse to evaluate proposed algorithm. Equal Error Rate(EER) of the proposed algorithm for this databaxse is 69.6 Practical results show that the proposed algorithm has an acceptable accuracy in comparison with the other presented methods.