کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Swin Wavelet Super Resolution

نویسندگان :

Zahra Moammeri¹ Ahmad Mahmoudi-Aznaveh²

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی

کلمات کلیدی :

super-resolution،transformers،wavelet transform،transform domain

چکیده :

Super Resolution has improved considerably by introducing novel approaches in the past few years. Transformers, with their abilities to capture long-range dependencies, are promising tools for super-resolution tasks. However, they suffer from the limitations of constrained patch areas. Since they process and analyze patches separately, achieving a wide and effective receptive field remains a challenge in these networks. In this work, we propose SwinWSR, a novel transformer-based model with a Spatial Wavelet block to capture more effective features through multi-scale wavelet transforms. We primarily focus on repetitive textures, such as those in the Urban100 dataset, which are particularly challenging to restore. Experimental results demonstrate that our model achieves competitive performance compared to existing methods, both quantitatively and qualitatively.