U-Net for creating training data for classification of DSPLs with poor resolution.

[PoohBear2]

All feedback is greatly appreciated!

To preface, DSPL imaging and classification is heavily restricted by poor resolution. The current problem is that computer generated images along with field observations of deep lensing, such as Einstein rings, are not only more abundant, but more clear, or at least at a higher resolution than the <5 DSPL’s that have been currently observed (by Dr. Nord). This is of particular problem for domain adaptation classification networks because this domain shift makes the target domain of DSPLs to be not very representative of the source domains that the DA’s are trained off of.

thus, to address the domain shift caused by poor resolution, I am proposing to create a U-net that is able to predict the residual of an image, and thus, use it to enhance the clearity of existing DSPLs which would in turn make the effects of domain shift less significant.

The fundamental reason behind the effectiveness of good image resolution enhancement machine learning modules is that they are able to predict the residual between a clear and unclear image very well. How it works is this:

1. To train, we use the clear deep field images and make unclear images for training by performing cubic interpolation with shrinkage and expansion on the clear images.
2. The model will then use this unclear image to predict its residual from the clear image from which it was created. To measure the accuracy of the network, we simply combine the residual with the blurry picture and measure the differences between the output and the original, sharp image with differences in pixel intensity
3. Once trained, it can be applied to the DSPLs so that classification can occur easier because the DA algorithms will still be trained on clear images, and now, with heightened resolution, the domain shift problem is no longer as big of an issue.

What do you guys think?