US20250278820
2025-09-04
Physics
G06T5/60
The patent application details a novel method for training an image restoration model that leverages diffusion models for unsupervised learning. The approach involves pre-training the model using a synthetic dataset, followed by refining it with real degraded images. This process generates initial restored images, which are then processed through a denoising diffusion model to create pseudo-target images. The difference between these initial and pseudo-target images is used to calculate a training loss, which is then utilized to adjust the model parameters, resulting in a trained image restoration model.
Image restoration is crucial in computational photography, aiming to enhance low-quality images affected by blurring, noise, or compression artifacts. Blind image restoration, where degradation processes are unknown, poses significant challenges. Traditional methods often rely on supervised learning with paired datasets, but these may not generalize well to unanticipated degradations and require extensive datasets. Diffusion models offer an alternative with robust modeling capabilities but are computationally demanding during inference.
The proposed method addresses these challenges by using unsupervised training that does not depend on ground-truth target images or knowledge of specific degradations. It utilizes a two-step process: pre-training with synthetic data followed by refinement using real degraded images processed through diffusion models. This approach reduces complexity compared to full diffusion models and aims to maintain both fidelity and perceptual quality of restored images.
Key steps include generating pseudo-target images via a forward diffusion and reverse denoising process applied to initial restored images. The reverse process involves constraints on high-frequency components during certain steps, while other steps remain unconstrained. This selective denoising enhances the restoration model's ability to handle diverse degradations effectively without excessive computational overhead.
The system can be implemented on electronic devices equipped with processors and memory to execute the described training method. It involves storing instructions for pre-training, generating pseudo-targets, calculating losses, and adjusting model parameters. Additionally, the system can restore degraded images using the trained model. The approach offers a scalable solution for real-world applications where large paired datasets are unavailable.