Defocus-enhanced technique for real-world scenarios using generative models

In recent years, significant progress has been made in deep learning-based image deblurring. These approaches utilize deep neural networks to learn the map between blurry and clear images or jointly learn the blurry kernel and clear image. They have demonstrated effectiveness in enhancing image quality, preserving details, and handling various types and degrees of blur. The objective of this study is to develop a defocus enhancement technique for real-world scenarios using score-based generative models. Stochastic differential equations (SDE) are employed to gradually introduce noise, thereby smoothing the data distribution towards a known prior distribution. The Score-Matching Langevin Dynamics (SMLD) model estimates the score for each noise scale, while Diffusion Models (DDPM) train the target model for score computation. This process constructs a score-based model capable of reversing the SDE over time. A predictor-corrector framework corrects the evolution of the reverse-time SDE, and the prior distribution is transformed back to the data distribution by removing the noise. By leveraging score-based generative models, accurate score estimation and sample generation are achieved using neural networks and numerical SDE solvers. This technique effectively restores clarity and details in defocused images, thereby enhancing overall image quality.