The mind's eye: reconstructing noise corrupted objects, extracting secondary structure and figure ground separation

摘要

The SLME network is a massively parallel recurrent, iterative, multiple constraint satisfaction neural network. It is capable of learning to solve problems including segmentation and vision tasks. It develops its own knowledge and feature set during iterative training. It consists of a retina of cells that learn their behavior and function during training. The cells are local processors that achieve consistency in an iterative manner using knowledge gained through learning, combined with vertical and horizontal communication of state between cells of neighboring regions. The SLME network can be used to extract structure from gray scale images of a naturally variable population for which it has been trained. Examples included are reconstruction and extraction of basic primary and secondary structure of noise corrupted objects and patterns, filling in objects, figure ground separation, and extraction of center axes of noise corrupted blobs.