Constraint satisfaction neural networks for image segmentation

Abstract

Image segmentation is a process to divide an image into segments with uniform and homogeneous attributes such as graytone or texture. An image segmentation problem can be casted as a Constraint Satisfaction Problem (CSP) by interpreting the process as one of assigning labels to pixels subject to certain spatial constraints. A class of Constraint Satisfaction Neural Networks (CSNNs), different from the conventional algorithms, is proposed for image segmentation. In the network, each neuron represents one possible label of an object in a CSP and the interconnections between the neurons constitutes the constraints. In the context of image segmentation, each pixel in an n × n image can be considered as an object, i.e. there are n 2 objects in the CSP. Suppose that each object is to be assigned one of m labels. Then, the CSNN consists of n × n × m neurons which can be conceived as a three-dimensional (3D) array. The connections and the topology of the CSNN are used to represent the constraints in a CSP. The initial condition for this network is set up by Kohonen's self-organizing feature map. The mechanism of the CSNN is to find a solution that satisfies all the constraints in order to achieve a global consistency. The final solution outlines segmented areas and simultaneously satisfies the given constraints. From our extensive experiments, the results show that this CSNN method is a very promising approach for image segmentation. Due to its network structure, it lends itself admirably to parallel implementation and is potentially faster than conventional image segmentation algorithms.