This file provides the dataset and code for Hierarchical Synthesis Convolutional Neural Networks(HSCNN) method. We proposed a novel method: HSCNN, to split the training data on domains of potential biases. This separation pre-processing allows for a guided learning, where physical insights of reconstruction aberration and noise can be integrated into the learning process to avoid overfitting. A hierarchical synthesis network is employed which is more adaptive to data pool with multiple sampling biases. The synthesis stages of the network enable the rebalance of the data against different sampling biases one by one. As a result, the learning scheme is more robust against sampling bias and aberrations introduced in the forward modeling. This method is very efficient that works with multi-band information without introducing dense inter-band connections. Much less computational efforts are required compared to dense connected DNN approaches.
Robust X-ray Sparse-view Phase Tomography via Hierarchical Synthesis Convolutional Neural Networks
- In revision process: []
- Implemented in Matlab 2018a with deep learning toolbox
Here is the link to download all training dataset: [https://www.dropbox.com/sh/uqsd5fl3ho22ucu/AAAdq4RAkqmG5ewElVt2KoNPa?dl=0]