Early vision image analyses using ICA in unsupervised learning ANN
Mohd Ruhul Ameen, Harold H Szu · 2003
Major problems in early vision are the edge feature extraction and segmentation of objects in order to recognize them separately. The paper presents a systematic methodology to the image analyses based on a breakthrough in unsupervised artificial neural networks by several groups in Europe, US and Japan, as motivated by blind source separation studies. In the unsupervised learning algorithm the features can be learned without teachers at the maximum entropy output of the artificial neural networks. The unsupervised algorithm may be paraphrased as "squeezing noise out and, thus without teachers, the feature edges are kept within": which incidentally reduces the redundancy and becomes pseudo-orthogonal to one another, i.e. ICA.