Data Dependent Weight Initialization in the Hopfield Neural Network Classifier: Application to Natural Colour Images
Sammouda, Rachid . 2010
The initial weight matrix to be used in the unsupervised hopfield neural network (HNN) image based classification or segmentation has a strong influence in the quality of the solution obtained after a specified time or iteration number given to the network to find the best solution and converge. An inadequate initial random matrix may cause the classifier to start with a big sum of errors in its first distribution of pixels among a pre-decided number of clusters, and get stack in a poor local minimum far from the global optima. In this paper, we present a new approach to initialize the weights of HNN classifier, dependant on the data to be classified by the network. The assignment of pixels to clusters is not purely random but restricted to a finite set of choices related to the information given to the network about the pixel. It is found that the performance can be improved with respect to the random initialization scheme.
Presents contributions to improve a previously published approach for the segmentation of magnetic resonance images of the human brain, based on an unsupervised Hopfield neural network.
The paper presents a method for automatic segmentation of sputum cells color images, to develop an efficient algorithm for lung cancer diagnosis based on a Hopfield neural network.