ニューラルネットワークの分類精度に関する基礎的検討とその切羽画像解析への応用

概要

論文の詳細を見る
Classification of geoscience data is an important tool for exploration of natural resources and interpretation of geophysical prospecting data. The minimum Euclidean distance method (MED) and the maximum likelihood method (ML) are typical methods in the supervised classification. These methods require that the boundary shape between two groups is linear and the frequency distribution of attribute data in each group is approximated by the multi-dimensional Gaussian distribution. A neural network (NN) can also be used in the classification and does not require the above conditions. The high accuracy of classification by NN over MED and ML was demonstrated in this paper using the three simple models with different shapes of group boundary. NN was applied to the classification of color image of excavated faces in the epithermal gold deposit in southwest Japan. In this mine, the ore grade is generally related to the veins color, e. g., the white zone is rich in gold and silver, whereas the brown zone being strongly altered is low grade zone. Training samples were selected automatically through the self-organizing feature map. Using these training samples and the NN of multi-perceptron type, the color images of veins were classified into three parts：high, middle and low grade parts. Because the classification results were revealed to be appropriate by the comparison with the maps created by geologists, the effectiveness of NN for the classification of complicated data with several attributes was proved.
社団法人資源・素材学会の論文
1998-03-25

ニューラルネットワークの分類精度に関する基礎的検討とその切羽画像解析への応用

スポンサーリンク

概要

著者

関連論文

スポンサーリンク