Massively parallel networks for edge localization and contour integration—adaptable relaxation approach

Abstract

The paper describes a new adaptive neural network for edge-based pattern extraction problems. In the model, each neuron represents an edge with continuous state variables describing its location and orientation. The post-synaptic distribution is dependent on the state variables. Each neuron adjusts its state to increase its membrane potential, which results in highly adaptive dynamics of the synaptic weight distribution. The network allocates multiple neurons with different orientation modes for each edge. The strategy allows accurate modeling of multi-modal distributions at key-points such as corners and junctions. As a result, the network delineates edges at sub-pixel accuracy while preserving key-points. It is also capable of processing a sequence of images and following moving objects. The network is extended for contour integration and key-point detection tasks. The paper presents experiments conducted on both synthetic and non-synthetic data to demonstrate the effectiveness of the technique.