| 摘要 |
PURPOSE:To calculate the most satisfactory correspondence with the small amount of calculation by using a hop field type network. CONSTITUTION:Two dot sets, namely, a model feature point qi(i=1, n) and a feature point pj(j=1, n) as the border line of an object are considered. Next, on a correspondent line between the model feature point qi and each input feature point qj, a neuron Nij is arranged. An output Vij of the neuron Nij is equipped with values from 0 to 1. Then, it is analyzed that 'the feature points qi and qj are not made correspondent' in the case of 0 and that 'the feature points qi and qj are made correspondent' in the case of 1. At such a time, total energy E of a neural network is defined as the sum of energy E1, which is determined so as to obtain a minimum value when each model feature point qi is made correspondent to the input feature point qj by one- to-one, and energy determined so that correspondence error can be made mini mum in the meaning of minimum square at the time of corresponding most suitably. Thus, by calculating the combination of the Vij so that the both E1 and E2 can be made minimum, the combination is the optimum combination of resolutions. |
