Figure-Ground Separation for Dynamic Image with Ginzburg-Landau Equation by Use of Local Clustering

Abstract

In many conventional image recognition methods, distances among pixels are rarely taken into consideration explicitly because images are processed in the form of vectors. Then the nunber of connection among pixels is in direct proportion to that squared of pixels. It is thought that this problem may be got over by reducing connections among distant pixels, for correlations among neighboring pixels are larger than among distant pixels in general images. This idea is realized by methods based on the autonomous decentralized system. This research is the first try to solve the problem of figure-ground separation for dynamic image with this method. The separation mentioned above is to extract the part (figure) at where motion velocity is different from the background (ground) in the dynamic image. This process consists of two phases. First we execute provisional separation using spatial differences of optical flow calculated locally at each pixel. Unfortunately, this provisional separation by local operation may include certain classification error. Then next we correct them by dynamical interactions among neighboring pixels. Now, problem of explosion in number of connection among pixels never occur because this recognition system is realized with Ginzburg-Landau equation which can be calculated by local operations. This recognition system is superior in error ratio to the method based on Bayes decision which uses global information and is the best method of linear separation. Furthermore, this system can execute the separation of dynamic image made of random dot pattern which have been thought to be difficult to separate.

Parallel processing of figure-ground separation in random dot kinematogram with optical flow discretization

“…Regarding visual process of RDK, Ueyama et al [4], [5] have succeeded a figure-ground separation by describing the recognition process with the bistable Ginzburg-Landau equation. Further, Okura et al [6] proposed a method for restoring the 3D shape of objects that are separated from RDK as figures.…”

“…Yuasa et al has described a visual processing algorithm of auto-associative memories [2], binocular stereograms [3] and figure-ground separations in animation [4] using a special class of differential equations, reactiondiffusion equations. In this paper, we also consider a method for the figure-ground separation in random dot animations using the reaction diffusion equations, where we aim at distinguishing multiple objects (figures) in the image.…”