The problems arising from nonorthonormality of the stored images in an optical resonator neural network are discussed, and a solution is given whereby the eigenmodes of the resonator can be found. This approach is implemented in computer simulations of the resonator. A number of diverse and distorted input images are successfully identified. Results for individual examples are presented and type I (in-class discrimination) and II (out-of-class discrimination) false alarm rates for this configuration are obtained. We show that by using the information on the convergence rate of an iterative scheme, the performance of the system, as compared to that of a correlation filter, can be improved.