Perceptron is a binary classification model in the field of pattern recognition， which has the advantages of simplicity， linearity and high computational efficiency， and it is also the basis of many classifiers. However， the perceptron cannot express complex nonlinear mapping and is difficult to process nonlinear data. Aiming at the inherent defects of perceptron and combining the characteristics of granular computing， we propose a new perceptron classification model—Granular perceptron. According to the theory of granular computing， the granulation of samples on single feature forms granules， and the granulation on multiple features constructs a granular vector. Further， the granular perceptron model is defined， the granular perceptron strategy is designed， and the granular perceptron learning algorithm is proposed. In order to solve the optimal solution of the proposed model， the derivative form of the granular loss function is proved， and its gradient descent algorithm is designed. Finally， some experiments are carried out to compare on the convergence speed， nonlinear ability and classification accuracy. The results show that the proposed model has the ability of fast convergence and nonlinear data processing.