衝撃貫通画像処理のためのニューラルネットワークの判別能力改善法

概要

論文の詳細を見る
The estimation method of the material's characteristics from the impact perforation images has been studied in the material engineering field. In this method, the characteristic of the material specimen is estimated from the steel ball speed measured in a high-speed photograph. However, the observation of behavior in the steel ball is often difficult because of the fragment of the specimen. We have studied to use a multilayer neural network to localize the steel ball in the impact perforation images. In the image distinction by the neural network, the data are distinguished by the input-output relation obtained beforehand by the training data. In general, other data are used for training and actual distinction. The neural network overcomes the differences between the training data and distinction data by the robustness. To improve the distinction accuracy, it is necessary to clarify the feature of the training data and to decrease the difference between the training data and the data for distinction. In this study, we introduce the preprocessing by the gamma correction to clarify the feature of the input images. Moreover, we introduce the incremental training method to decrease the difference between the training data and the data for distinction. We confirm the effectiveness of these two methods by the computer simulation which uses an actual impact perforration images.
拓殖大学の論文