Convolutional neural networks for fast myopic deconvolution of AO observations

CONVOLUTIONAL NEURAL NETWORKS FOR FAST MYOPIC DECONVOLUTION OF AO OBSERVATIONS Pierre Vermot (LIRA, Paris Observatory, PSL) AO4ELT8 –28/10/2025 R. Petrov (ANR AGN_MELBA), D. Gratadour (PEPR Origins), Y. Clénet, A. Berdeu, B. Besnard, ... TALK PLAN - WORKFLOW CNN for deconvolution 2 I. Sky model - Targets image generation II. Instrumental model – Observations generation III. CNN architecture and training IV. Testing V. Applications VI. Conclusions I. SKY MODEL CNN for deconvolution •Combination of Fourier Random Fields: oPowerlaw spectrum + random phase oFourier Transform oPoint sources 3 Image generation algorithm II. INSTRUMENTAL MODEL – OBSERVATIONS GENERATION CNN for deconvolution 1. Short exposure PSFs •Residual atmospheric phase: oCombination of Zernike modes oRandom Strehl ratio •Pupil •FFT 4 II. INSTRUMENTAL MODEL – OBSERVATIONS GENERATION CNN for deconvolution 3. CCD frame generation oConvolution of Source image with LE PSF oPhoton noise, read-out noise 5 4. Full observation oRepeat steps (1, 2, 3) to generate data cubes with K images of the same object 2. Long exposure PSFs Combination of N short exposure PSFs CNN ARCHITECTURE AND TRAINING CNN for deconvolution U-NET model •Fully convolutional auto-encoder •Skip connections •Widespread architecture for image segmentation, deconvolution/super-resolution, computer vision 6 CNN ARCHITECTURE AND TRAINING CNN for deconvolution U-NET training •Trained to reconstruct the original image (output) based on the convolved datacube (input) •Minimizes Mean Squared Error •1M examples, up to 50 epochs •Training ~ several hours •Inference ~ 10-3 s 7 CNN for deconvolution 8 CNN for deconvolution 9 COMPARISON WITH CLASSICAL DECONVOLUTION ALGORITHMS Evaluation on realistic data •Real IR astronomical images: rebinned 2MASS K-band •Experimental PSF: NaCo bright star •Several metrics: oSquared residuals oR2 oUniversal Image Quality Index oSeveral algorithms: oRichardson-Lucy oMAP oWiener CNN for deconvolution 16 COMPARISON WITH CLASSICAL DECONVOLUTION ALGORITHMS Evaluation on the validation dataset •Several metrics: oSquared residuals oR2 oUniversal Image Quality Index •Several algorithms: oRichardson-Lucy oMAP oWiener