Гібридна модель реконструкції зображень із неповних вимірювань на основі розгорнутої оптимізації та дифузійного апріорного моделювання

Abstract

The thesis addresses image reconstruction from incomplete and noisy measurements in computational imaging. A hybrid method that combines unrolled optimization, explicit data consistency and a diffusion prior is proposed. The model separates the physical acquisition operator from the learned image prior and thus supports adaptation to different degradation scenarios. This design is intended to reduce oversmoothing typical of classical regularization and to limit hallucinated details produced by purely generative models. Special attention is paid to balancing fidelity to measured data with perceptual plausibility. For evaluation, it is advisable to combine PSNR and SSIM with LPIPS to capture both pixel-wise and perceptual quality. The approach is relevant to MRI, CT, microscopy and super-resolution tasks and may improve robustness across domains