深層学習用活性化関数によるニューラルネットワーク選点法トモグラフィの改良

抄録

Neural Network Collocation Method (NNCM) is a kind of model fitting method and is an effective imaging technique for application plasma tomographic imaging field that has limited to the number of views. The reconstructed image is represented as nonlinear combination of the basis functions by the activation functions of NNCM. NNCM could fit models based on few spatial location data because every basis function cover whole reconstruction region and are trained based on error back propagation algorithm. Therefore, the image could be effectively reconstructed even from few-view poor projection data. However, application of NNCM to practical imaging has an issue that NNCM requires many training epochs even to reconstruct a small-scale image. The main reason of the issue is attributed to that the small gradient value of sigmoid activation functions let the training speed slow.

The paper proposes an improvement of NNCM by replacing the activation functions of hidden layers of NNCM with deep learning type ones which have larger gradient values, ReLU and tanh, to increase training speed of NNCM. Numerical simulation results show the effectiveness of the proposed method.