Learning of view-invariant pattern recognizer with temporal context

摘要

Radial basis function (RBF) neural networks have been used for view-invariant recognition of three-dimensional objects on the basis of combination of some prototypical two-dimensional views of objects. Supervised learning algorithms have been used for optimizing their parameters with the input of image set of some objects from several views. We modify its structure slightly and develop an unsupervised learning algorithm with the input of a time sequence of continuously variant view images. A feedback path is added to RBF networks for utilizing class-membership outputs at a previous time as a prior information at the current time. The network is trained unsupervisedly by a maximum likelihood scheme with the input of time-variant view images of some objects for developing view-invariant recognition capability. Robustification of data distribution enables the network to reject outlier data and to extract an object expected from the recognition at previous times.