Image Segmentation Based on Consensus Voting

Abstract

This paper presents a new approach called consensus voting neural network (CVNN) which aims to perform fast image segmentation for grey images. A learning algorithm based on the principle of vote-to-consensus is developed to train CVNN. The essence of CVNN is the iterative interaction between the target neuron and its neighboring pixels, and the range of neighborhood is defined by the running mask. The neighboring neurons surrounding the target neuron collaboratively determine the label for the target neuron by "casting" their respective labels. Due to its simplicity in the updating strategy that solely employs discrete increment value in the ballot-counter, training CVNN is quite efficient.