Relaxation neural network for nonorthogonal image transforms

Several image-processing problems require finding representations for 2-D signals in terms of expansion functions which, in general, may be either orthogonal nor complete. Finding the desired set of coefficients or feature descriptors in general can be difficult, both because of the nonorthogonality of the representation and because of the high dimensionality of (say) a 512*512 image. The present approach formulates the calculation of such coefficients as an optimization problem, which a three-layered relaxation network then solves. Examples of applications which are illustrated with nonorthogonal (yet complete) 2-D ´Gabor´ transforms include: (1) image compression to below 1.0 b/pixel, and (2) textural image segmentation based on the clustering of the coefficients found by the relaxation network.<>