In this paper, we address non-linearities in images to approach flexible templates. Templates reflect objects of a single class and are extended to have the special ability to cover the variations present about the object. An auto-associative neural network learns these variations from examples. We consider images to be related to an artificial retina where the appearance of observed objects is represented. Prom this point of view, non-linear grey-level changes are the consequences of global and local variations of the object. Image variation is considered in a high-dimensional image space. Thus, varying objects from the same class leave a manifold in the image space, which is modeled by the introduced network.
展开▼
