Self - Organizing Maps and Computer Vision

Publisher Summary This chapter emphasizes the role of unsupervised learning and self-organization in some stages of a computer vision system. Neural networks can be used in an efficient manner at the intermediate level image processing tasks, especially in global feature extraction. The term global means that features either are not localized to small pixel neighborhoods spatially or combine different resolutions or otherwise represents the combinations of the outputs of primary feature detectors. These combinations can be coded by a self-organizing feature map to obtain compressed representations that can be used at later stages. The chapter discusses the use of the self-organizing feature map in a typical intermediate level computer vision task, texture feature extraction, and texture image segmentation,. It presents an example to show that in curve segment detection, sometimes the central characteristics of neural network learning algorithms, particularly the random sampling from the input space and the mapping of the high-dimensional input space to individual neural units, can give valuable insights in developing efficient nonneural image processing algorithms. The conclusion of the randomized Hough transform algorithm is that it has several advantages over the conventional Hough transform that is in the way the accumulator array is replaced by a dynamical tree, in which new nodes can be created at will, and the parameter resolution and parameter scope is arbitrarily high.