Plots that we need to evaluate the model and we will eventually include in the paper.
Preliminary tests
Here we train and test only using one exposure. The number of test stars will be fixed to a high star density (~2400 stars in the exposure) and won't contain noise, and the number of train stars will change depending on the experiences.
We would like to test 3 situations of training stars:
[x] [Situation A] Having a high star density: ~60 stars/CCD.
[x] [Situation B] Having a normal star density: ~25 stars/CCD.
[x] [Situation C] Having a low star density: ~15 stars/CCD.
Now, for each of these situations we need a dataset that contains test stars without noise and train stars with and without noise. We would like to test:
[x] [Test 1] Classic MCCD without denoising trained with noiseless stars.
[x] [Test 2] Classic MCCD without denoising trained with noisy stars.
[ ] [Test 3] Classic MCCD with wavelet denoising (K=3.) trained with noisy stars.
[x] [Test 4] Deep MCCD (Learnlets) trained with noisy stars. Using uniquely trained Learnlet.
[ ] [Test 5] Deep MCCD (U-net) trained with noisy stars. Using uniquely trained U-net.
[ ] [Test 6] Deep MCCD (Learnlets) trained with noisy stars. Using one Learnlet for local components and one Learnlet for global components.
[ ] [Test 7] Deep MCCD (U-net) trained with noisy stars. Using one U-net for local components and one U-net for global components.
For each test we would like to use as metrics:
Pixel RMSE.
e1, e2 and R2 RMSE measured with Galsim's HSM module. This should be done for the noiseless test stars.
e1, e2 and R2 variance.
First the results can be presented as a table, but we would want to have different plots for the different metrics (pixel RMSE, e1 and e2, and R2) that show the evolution of the errors as a function of the star density for the different models.
Complementary tests
If we are using neural networks for the denoising, we would like to show that they are actually doing a good job at denoising. So, we would like to have some plots or a table maybe showing the pixel RMSE for the different networks on a standalone denoising task.
Plots that we need to evaluate the model and we will eventually include in the paper.
Preliminary tests
Here we train and test only using one exposure. The number of test stars will be fixed to a high star density (~2400 stars in the exposure) and won't contain noise, and the number of train stars will change depending on the experiences.
We would like to test 3 situations of training stars:
Now, for each of these situations we need a dataset that contains test stars without noise and train stars with and without noise. We would like to test:
For each test we would like to use as metrics:
First the results can be presented as a table, but we would want to have different plots for the different metrics (pixel RMSE, e1 and e2, and R2) that show the evolution of the errors as a function of the star density for the different models.
Complementary tests
If we are using neural networks for the denoising, we would like to show that they are actually doing a good job at denoising. So, we would like to have some plots or a table maybe showing the pixel RMSE for the different networks on a standalone denoising task.