本計畫利用類神經網路結合深度學習法，重建電磁波各向異性物體之成像。吾人發射TM(transverse magnetic)及TE(transverse electric)電磁波照射各向異性物體，並接收來自物體之散射場，由此散射場先算出主要等效電流(dominant equalvent current)，再算出各向異性物體之介電係數，算出之介電係數和真正的介電係數會有少許誤差，吾人再利用人工智慧中的類神經網路，如多層感知機(Multilayer perceptron, MLP)、輻狀基底函數(Radial Basis Function)和卷積神經網路(Convolutional Neural Networks)。於類神經網路中加入深度學習法，重建出最後的介電係數，藉此降低與真正的介電係數之差距。先研究針對雙軸性物體，將入射波分為兩部分，第一部分發射TM波，大量測試深度學習中的各種方法及參數、卷積神經網路架構，提出一個適用於電磁波成像之組合，並重建TM影像(z方向之介電係數)。第二部分利用TE波入射，利用散射場、深度學習法中適合之方法，並更深入研究卷積神經網路架構，重建TE之影像(x,y方向之介電係數)，藉此降低與真正的介電係數之差距，TM和TE影像分別對應，雙軸物體之介電係數。最後推廣上述兩種方法重建各向異性(Anisotropic)之物體。
This proposal is using neural network and deep learning method to reconstruct electromagnetic wave imaging of anisotropic objects. We transmit the electromagnetic wave to the anisotropic objects and record the scattered field. By TM and TE waves illumination, the dominant equivalent currents can be calculated. According to the dominant currents, we can compute the approximate permittivity tensors. Note that this approximate permittivity tensors will be a little different with the exact permittivity tensors, since only dominant currents are used. As a result, we use the neural networks (such as multi-layer perceptron, radial basis function and convolution neural networks) in the artificial intelligence to reconstruct the exact permittivity tensors by deep learning. In the first part, we will focus on the TM (Transverse magnetic) wave to reconstruct the permittivity in the z direction. In the second part, we will focus on the TE (Transverse electric) wave to reconstruct the permittivity in the x and y direction. Finally, we investigate the reconstruction of anisotropic objects. We will improve the reconstructed image by applying the dominant equivalent current and deep learning. Suitable neural networks to reconstruct the permittivity tensors will be proposed.
