A complex nonlinear exponential autoregressive CNEAR model for invariant feature extraction is developed for recognizing arbitrary shapes on a plane. A neural network is used to calculate the CNEAR coefficients. The coefficients which constitute the feature set are proven to be invariant to boundary transformations such as translation rotation scale and choice of starting point in tracing the boundary. The feature set is then used as the input to a complex multilayer perceptron C-MLP network for learning and classification. Experimental results show that complicated shapes can be accurately recognized even with the low-order model and that the classification method has good fault tolerance when noise is present.