Presentation by Maxime Paillassa

[sfarrens]

Robust detection of astronomical sources using convolutional neural networks

Speaker: Maxime Paillassa
Date: 29/04/2021
Slides: PDF

[andrevitorelli]

Quick notes:

cnn for robust detection

Maxime Paillassa

• astronomers usually work on catalogues
• problems: deblending, nearby stars
• robust: identify contaminants (image defects)
• Context: Cosmic-DANCe (IMF of stars)
• Heterogenous data, with proper motions...
• Context2: Euclid
• Huge data volume space/ground images (matching bc euclid photometry not enough for z-ph)
• Sextractor++
• current: several steps: bkg subtraction, matched filter, peaksearch/thresh/deblending
• matched: correlation signal/pattern
• optimal linear filter -> mas SNR
• problems: sources have various scales, can overlap
• matched filter detects one pattern at a time -> problems
• use NN
• huge amounts of data available
• image -> semantic segmentation -> instance(-aware semantic) segmentation
• Mask R-CNN: boxes, rescaling implies interpolations, no deblending, needs clear boundaries
• multiscale footprint segmentation
• brightest pixels, pixels above a fraction, compromise, configurable

Alexandre:

• def is relative to max intensity, no relation to the noise?

Maxime:

• yes, on noise-free, isolated sources (training) -> SkyMaker (Bertin 2006)

• assign a scale to sources

• galaxies (training): cosmo sims, stars SkyMaker

• train images are contaminated to train robustness

Sam:

• statistics over SX?
  Maxime:
• Over test sets.

Bastien:

• plan to compare ground/space telescope?
  Maxime:
• Good point HST images as ground truth
  Bastien:
• multiband detection? how to manage footprint different detection
  Alexandre:
• as per definition, I think there wouldn't be much of a difference between filters

Axel:

• what is the actual output: pos of peak? more?
  Maxime:
• footprint map probability of pixel being part of detection
  Axel:
• how to combine scales together?
  Maxime:
• for now not much confusion. postprocessing
  Axel:
• here, for example (slide 19): how do you plot the objects?
  Maxime:
• combination of all scales
  Axel:
• would we then not need masking?
  Maxime:
• that's one of point of this, yes
  Sam:
• how do you trust when it's sensitive to the detection
  Maxime:
• this is tricky, more simulations to quantify
  Axel:
• q related: more sources CNN, properties of that sources: are they low SNR? gain from deblending?
  Maxime:
• cnn can really find fainter sources, apparently
  Alexandre:
• what is sextractor here, because you can tune it in many ways
  Maxime:
• this example, gaussian kernel (5x5), 1.5 x sigma, no hot/cold approach
  Alexandre/Maxime:
• people will say that they can run sextractor better, the NN is not tunned.

Alexandre:

• speed?
  Maxime:

• 1mpixel/sec + preprocessing
  Sam:

• important is if you can process your large area

Alexandre/Maxime:

• train on noiseless, and then in noisy? Noiseless footprint, but training on noisy

Alexandre:

• NN has to be trained on different noises?
  Maxime:
• now maybe the range is not enough, but we could expand

Alexandre:

• could compute completeness vs snr
• it seems that footprints are smaller
• Deblending? multiple scale approach is good, but also obj can have star forming regions...

Axel:

• bkg: how sensitive if bkg is not well subtracted? can the network avoid bkg subtraction?
  Maxime:
• i think it's necessary. i didn't include very bad bkg, but yes variable... it should be able to detect in this conditions.
• the one that I use is same as sx.
  Axel:
• last q: varying psf not smoothly, is that a problem, like, in coadds -> variation of the size of stars. would that be a problem to the CNN
  Maxime:
• not tested, future simluation.