Color Image Segmentation Based on Lattice Auto-associative Memories

This manuscript introduces a new technique for autonomous segmentation of color images in Red-Green-Blue (RGB) space that makes use of lattice auto-associative memories (LAAMs). LAAMs are artificial neural networks able to store a finite set X of pattern vectors and retrieve them almost correctly when a noisy or incomplete input is presented. Two dual lattice auto-associative memories have been established, the min memory WXX and the max memory MXX whose column vectors, scaled appropriately, are used to determine a tetrahedron enclosing a subset of X . Specifically, the column vectors of each memory will correspond to the most saturated color pixels. Thus, from the perspective of convex set geometry, the scaled column vectors of a LAAM are extreme points of a convex subset of X (tetrahedron), and any other pixel can be considered a linear mixture of these points. Unsupervised segmentation of a color image is then realized by unmixing pixels using the non-negative least square method. We provide illustrative examples to demonstrate the effectiveness of our method as well as segmentation results for some RGB color images.