Abstract:
A system with some variations is presented for extracting multiple 2D target structures from one or more grey-scale sensory images of objects. In particular, the objects may be naturally variable and in a cluttered scene. These systems learn about their target populations of images of objects and the associated target output structures during a training phase. The proposed models are based on neural networks and cellular local processing methods, resulting in a massively parallel computing paradigm. The system presented includes a novel architecture and training approach for training cellular machines to solve complex and poorly understood problems with ease. The models perform their function using an iterative short-time relaxation labelling scheme. Various examples are presented including basic "toy" examples for noise suppression, structure extraction, filling and object separation. Also, some examples of structure extraction of a population of naturally variable objects are presented.