This paper presents a new neural network structure from which features of input data can be extracted easily. The basic structure is a two-layer perceptron and each unit in the input layer has a fuzzy membership function. The input space is divided into fuzzy sub-spaces by the membership functions and the new type of neural network has an ability of nonlinear mapping. The new neural network can handle some deformations of input data with the fuzzy divisions. The structure of the new network is very simple and the features of input data are easily known from the connection weights of the learned network. The authors call the neural network Fuzzy Net. A basic experiment of signature recognition is done using the fuzzy net. The capability of the new network in feature extraction is verified.