Neural Networks for the Segmentation of Outdoor Images

N. W. Campbell, B.T. Thomas, Neural Networks for the Segmentation of Outdoor Images. International Conference on Engineering Applications of Neural Networks. ISBN 952-90-7517-0, pp. 343–346. June 1996. PDF, 273 Kbytes.


The SOFM is trained using colour and Gabor texture values extracted from a set of training images. It is shown that the SOFM is capable of dividing the total input space into a small number of meaningful clusters. These are then used to index each pixel in the image. Similarly indexed pixels correspond to regions having similar colour and texture properties, and hence a segmentation is available. One of the papers main contributions is to quantify the success of a SOFM for this segmentation task on a large set of outdoor scenes, and not a small number of simple, artificially textured images. The high dimensionality of the input data is successfully reduced using this method in such a way as to allow segmentation. Practical considerations discussed include how the size of SOFM, the training regime required and the input dimensionality affect the segmentation quality.

Bibtex entry.

Contact details

Publication Admin