Search

The goal of this work is to study a preprocessing method for the data measured by a two-dimensional optoacoustic tomograph in order to reduce or eliminate artifacts introduced by the low number of detectors in the experimental setup and their limited bandwidth. A generative adversarial deep neural network was used to accomplish this task and its performance was compared with a reference U-Net neural network. In most of the test cases carried out, a slight improvement was found by applying the proposed network when measuring the Pearson correlation and the peak signal noise ratio between the reconstructed image product of the data processed by the model and the high-resolution reference image.

We first present a detailed review of two popular classes of model based synchrophasor estimators. They are the discrete Fourier transform (DFT) and the Taylor-Fourier transform (TFT) class estimators, which arise from constant or polynomial phasor models, respectively. Limitations of these techniques, arising from the models themselves, are exposed. These limitations make the original DFT and TFT approaches unfitted for situations where the phasor can not be properly described in such simple terms. This issue can be addressed with the recently introduced convex semi-infinite programming (CSIP) based estimators, which are also reviewed. In particular, we present the constraints associated with the tests which are most critical for the model based methods, showing how to precisely control the estimators performance in those cases. Finally, we show that the DFT and TFT estimators are particular instances of the CSIP estimator, so that there exists a connection between these two apparently different approaches. This opens the door for future analyses and developments.