Wykaz obszarów badawczych związanych z tagiem Deep-neural-networks-for-image-reconstruction-in-electrical-capacitance-tomography_1:
| # | Obszar badawczy | Dziedzina naukowa |
|---|---|---|
| 1 |
Capacitance measurements will be estimated. The selection of training datasets that may be application-dependent will be performed. Real data generation will consist in dynamic image acquisition when test objects are moved and rotated in the tomographic probe. It will be required to equip the measuring station with appropriate mechanics. For obvious reasons, the number of test objects will be very limited compared to the numerical simulation. It will be necessary to use data augmentation methods. The measurements will be carried out using an EVT4 electrical capacitance tomography scanner which was entirely designed and built by our group. The architecture of the hardware is based on fast programmable devices and allows fast measurements with high signal-to-noise ratio using 32 channels. A comparative study on deep learning architectures in the context of image reconstruction will be performed. Networks that have already been used in electrical capacitance tomography or electrical impedance tomography will be considered as well as several popular deep learning architectures, such as convolutional neural networks, recurrent neural networks, long short-term memory/gated recurrent unit, self-organizing map, autoencoders and restricted Boltzman machine. Learning and testing of selected DNNs with different dataset configurations will be done. The different loss functions and different solvers will be used to optimize the networks. The images will be reconstructed using DNNs and a state-of-the-art algorithm (Levenberg-Marquardt) will be used as a base reference in comparative analysis. The results will be assessed using selected image quality norms. The conducted research should influence the shift of paradigms prevailing in the ECT environment. The use of deep networks may allow for increasing the spatial resolution of ECT scanners. The knowledge gained in the selection of a deep network, analysis of its architecture, and training techniques can be used in inverse problems occurring in other fields. The results of the research will be the subject of publications in renowned journals. All datasets, DNN architectures as well as software tools like ECTsim will be made available to the scientific community in public repositories.
|