ニューラルネットワークモデルを用いた緩和ラベリング法の提案

概要

論文の詳細を見る
The classical relaxation labeling method proposed by Rosenfeld, Hummel, and Zucker is very simple and widely used for solving various labeling problems. Its updating rule for labeling weights can be obtained from a transition rule of a linear Hopfield's analog neuron model by using a discrete time approximation. This neuron model, however, is not a natural neural network model because a time constant for each neuron should be modified at each time step according to a state of the system. In this paper, a more powerful relaxation method ("Neural Relaxation Labeling Method") based on a neural network model with a nonlinear output function is proposed and applied to problems of matching and recognizing patterns which consist of several straight line segments. Compatibility coefficients are calculated by a linear combination of several feature functions which represent disagreements between two pairs of primitives. Parameters including proportional coefficients with which the feature functions are multiplied can be learned by using a steepest descent method for decreasing the value of an error function. Our simulations indicate that the present method with the parameters above has achieved more correct matching ability and a higher level of recognition ratio than those by the classical relaxation method.
社団法人人工知能学会の論文
1991-11-01