A novel approach to object recognition and scene analysis based on neural network representation of visual schemas is described. Given an input scene, the VISOR system focuses attention successively at each component, and the schema representations cooperate and compete to match the inputs. The schema hierarchy is learned from examples through unsupervised adaptation and reinforcement learning. VISOR learns that some objects are more important than others in identifying the scene, and that the importance of spatial relations varies depending on the scene. As the inputs differ increasingly from the schemas, VISOR's recognition process is remarkably robust, and automatically generates a measure of confidence in the analysis.