Optimizing Loss Functions in Conditional GANs for Astronomical Image Deconvolution: A Comparative Study at Varied Epochs*

This study explores the optimization of loss functions in conditional Generative Adversarial Networks (cGANs), specifically the pix2pix model, for the deconvolution of astronomical images obtained from the Hubble Space Telescope. The objective is to mitigate the effects of atmospheric distortion, commonly referred to as the "seeing" effect. We systematically evaluated various combinations of Mean Absolute Error (MAE), Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Binary Cross-Entropy (BCE) with logits as loss functions for the generator. These were tested at different training epochs—10, 20, 30, 40, and 50. Our results indicate that a specific weighting of BCE with logits, MAE, and SSIM—1, 1, and 50, respectively—significantly enhances the performance, yielding higher Peak Signal-to-Noise Ratio (PSNR) and SSIM values post 30 epochs. In addition, we compared our cGAN-based method with conventional techniques, including Wiener deconvolution, Richardson-Lucy algorithm, and Total Variation (TV) deconvolution. Our proposed method demonstrated superior clarity and detail retention.