A neural network structure that learns feature extraction and classification operations simultaneously is described. The feature extraction operations are represented using generalized image algebra operations. Learning rules are described. Linear operations and nonlinear, hit-or-miss operations are used to perform handwritten digit recognition.