Similarity measure for pattern recognition

摘要

Disclosed is a method for obtaining an accurate detected value of a similarity between images or the like. Original standard and input pattern matrices, each having a feature quantity of an image as an element, are created (Sb1 and Sb3). A weighting vector having a value of a change rate of a kurtosis of a reference pattern vector composed of a reference shape as an element is created, and an original and weighted standard pattern matrix is calculated by a product-sum operation of an element value of the weighting vector and an element value of the original standard pattern matrix (Sb2). Then, an original and weighted input pattern matrix is calculated by a product-sum operation of the element value of the same weighting vector and an element value of the original input pattern matrix (Sb4), and a geometric distance value between the original standard pattern matrix and the original input pattern matrix is calculated as a cosine of an angle between the original and weighted standard pattern matrix and the original and weighted input pattern matrix (Sb5).