koc-lab / gfrft-unified Goto Github PK

Import median-filtering source codes from JFRT experiments.

Currently, the optimal-filtering-task/Comparison.m generates results by generating $k$-NN graph for a given $k$ value. Convert this structure to generate all results for a given array of $k$ values. With this update, the tabular printing also needs to be adjusted #7.

The "optimal filtering" experiments in JFRT paper correspond to the "denoising-task" in the initial submission of the Unified GFRFT paper. The optimal filtering experiments need to be imported for this paper.

• Import source code that works with ARMA and Median-Filtering frameworks (see #3 and #4).
• Simplify the joint time-vertex structure to graph signal structure.

Add a small script to print the obtained results in a LaTeX tabular format so that it would be possible to import .mat results directly to reports and manuscript.

Currently, the optimal filtering task creates and applies an optimal filter for each graph signal (each column) of the joint time-vertex (JTV) signal. For $N\times M$ JTV signal, the computational complexity is as follows:

1. Generating optimal filter for a graph signal $\boldsymbol{x}\in\mathbb{C}^N$ is $\mathcal{O}(N^4)$.
2. Generating the optimal filter for each graph signal in a JTV signal is $\mathcal{O}(MN^4)$.
3. Finally, going through a grid search for the fractional order with $N_a$ different orders sums up to $\mathcal{O}(MN^4 N_a)$.

Instead, one can estimate the auto- and cross-correlations of the true distribution, followed by generating the optimal filter based on this estimate and applying the same filter to all graph signals. This operation could reduce the overall obtained best SNR values, but would also reduce the complexity to $\mathcal{O}(N^4 N_a)$ which means ~ $300$ times faster experimentation since the datasets have between $100-600$ time instances.

Using a fixed $\sigma$ for Gaussian noise generates varying noise errors in the experiments, where switching to signal-to-noise ratio (SNR) based noise generation could fix this issue. In this context, providing SNR values after denoising and optimal filtering processes to compare before and after operations can be a better standard.

Import already generated ARMA Graph filtering codes obtained from Andreas Loukas's blog post. The time-varying version is implemented in JFRT, however it is not necessary for now.