To efficiently reduce dimensions of a feature space for the purpose of high precision, a high speed and memory saving in order to solve deterioration in identification rate in identification processing in a high-dimension feature space, increase in calculation amount, and increase in a use memory.
A method reduces dimensions of a feature space by selecting a partial space storing a major component of a quadratic function by leaning the quadratic function through a polynomial neural network by using a feature pattern group for generating a dictionary. An initial coefficient setting step 42 and a coefficient correction step 43 correct the coefficient by means of a method for gradient descent or a method for probabilistic gradient descent so that the value of a loss function becomes smaller when the quadratic function is used as an identification function. A base vector derivation step 44 derives an eigenvector of a matrix in a quadratic form of a quadratic term of the quadratic function and a coefficient of a linear term. Next, a projection matrix derivation step 45 selects one or more vectors to be a main component among the eigenvector and the coefficient vector and generates the partial space, as a new feature space, generated by the selected vector.
COPYRIGHT: (C)2010,JPO&INPIT
Ken Nagasaki
Hiroshi Shinjo
Shoichi Ishii
JP2007179307A | ||||
JP2006059284A |
Cheng-Lin Liu, et al.,Class-specific feature polynomial classifier for pattern classification and its application to handwritten numeral recognition,Pattern Recognition,2006年 4月,vol.39, issue.4,p.669-681
Next Patent: 灯火装置一体式バックミラー