Unsupervised Restoration

Unsupervised image restoration aims to recover high-quality images from degraded versions without relying on paired training data, a significant challenge addressed by leveraging generative adversarial networks (GANs) and optimal transport (OT) frameworks. Current research focuses on improving GAN-based methods, such as CycleGAN, through innovative loss functions and architectural modifications like incorporating self-collaboration strategies or sparsity-aware components to enhance performance and address limitations in handling specific degradations (e.g., stripes in infrared images). These advancements are crucial for various applications where paired data is scarce or expensive to obtain, leading to improved image quality in diverse fields like remote sensing and medical imaging.