Szkoła doktorska Politechniki Warszawskiej

Wyszukiwarka promotorów i obszarów badawczych

Wykaz obszarów badawczych związanych z tagiem Deep-neural-networks-for-image-reconstruction-in-electrical-capacitance-tomography:

# Obszar badawczy Dziedzina naukowa

Electrical capacitance tomography (ECT) allows to visualize a spatial distribution of electrical permittivity of objects. Images are reconstructed using capacitance values measured between electrodes surrounding the visualized volume. Due to a small number of electrodes (8 to 32), the number of measurements is much smaller than number of image elements which makes image reconstruction ill-posed and ill-conditioned. State-of-the-art algorithms like Levenberg-Marquardt are mostly iterative and nonlinear, i.e. they require many recalculations of the Jacobian matrix which constitutes relation between permittivity values and capacitances (sensitivity matrix). This is a very time-consuming task which makes real-time imaging very hard. It is necessary to acquire new knowledge that will overcome the existing limitations. The project will consider an alternative solution based on deep neural networks (DNNs).

The research planned in the project aims to investigate deep network architectures that will achieve not only comparable but significantly better results than classical methods. Various DNNs will be analyzed and tested. The behavior of the DNNs in tomography with the increased number of electrodes will be investigated (32-electrode sensor). We are curious if the use of neural networks will push the limits of image quality in ECT, especially image spatial resolution. Numerical modeling of electrical capacitance tomography sensors and test objects will be carried out using the ECTsim toolbox for MATLAB. The finite volume method (FVM) using non-uniform mesh will be applied to compute electrical field distributions and sensitivity matrices of sensors. The numerical modeling will allow to optimize the electrode configuration in the sensing probe. The arrangement of electrodes in one ring (32 electrodes) and in two rings of 16 electrodes each will be considered. Generation of datasets for learning, testing and validation will be performed. The training data will be generated numerically and using real measurements. The permittivity distribution in the probe will be randomized according to the assumed probability distributions describing the number of objects, their shape, location and the value of permittivity.