eric930711 / gsdd Goto Github PK

Single image super-resolution (SISR), especially in the real world, usually builds a large amount of LR-HR image pairs to learn representations that contain rich textural and structural information. However, relying on massive data for model training not only reduces training efficiency, but also causes heavy data storage burdens. In this paper, we attempt a pioneering study on dataset distillation (DD) for SISR problems to explore how data could be slimmed and compressed for the task. Unlike previous coreset selection methods which select a few typical examples directly from the original data, we remove the limitation that the selected data cannot be further edited, and propose to synthesize and optimize samples to preserve more task-useful representations. Concretely, by utilizing pre-trained GANs as a suitable approximation of realistic data distribution, we propose GSDD, which distills data in a latent generative space based on GAN-inversion techniques. By optimizing them to match with the practical data distribution in an informative feature space, the distilled data could then be synthesized. Experimental results demonstrate that when trained with our distilled data, GSDD can achieve comparable performance to the state-of-the-art (SOTA) SISR algorithms, while a nearly ×8 increase in training efficiency and a saving of almost 93.2% data storage space can be realized. Further experiments on challenging real-world data also demonstrate the promising generalization ability of GSDD.

Our experimental datasets include OST and DPED. The distilled OST image data (i.e., ICP=50, 100, 200, and 500) will be published soon.

This paper focuses on exploring the possibility of applying DD to benefit low-level CV tasks (especially for SISR). Overall, we propose a GSDD framework that optimizes and synthesizes training samples by the GAN-inversion manipulation in a latent generative space. Our optimization process is based on a distribution- matching data condensation strategy, so that the SISR network can have comparable performance to models trained on the original dataset. We further improve the approach by proposing a distillation loss with the regularization term R. Finally, we demonstrate its effectiveness via extensive experiments. We achieve a competitive performance compared to SOTA SISR solutions under the premise of a nearly ×8 increase in training efficiency and a saving of almost 93.2% data storage space. We wish that the study opens a new perspective in SR research and expect to see it facilitates more practical applications in the future.