Glass pattern recognition with neural-networks

The visual system analyses many of its perceptual problems in a two-stage model which consists of a local stage, a global stage, and a connection between them through feedback. This two stage model has been used for classification of images which are characterized uniquely by the orientation feature. Such images are made out of random-dot textures with local connections between adjacent dots. An example is the glass patterns which are created by superimposition of two identical random-dot patterns, while inducing a transformation such as rotation, expansion, and translation of one of the patterns, thus creating a global percept of structured circles, stars or straight lines, respectively. In the local stage an energy function is minimized via the Hopfield model, to achieve oriented dipoles between corresponding dots. In the global stage the Back-Propagation model for the learning procedure is used with a new energy function to achieve maximum likelihood decision rule by ratio of conditional probabilities, for the discrimination between images. Possible feedback between the stages is checked. It has been shown that the neural model can learn to distinguish between different global organization effects through the orientation fields.