Deep learning methods for blind super resolution using self-attention transformers and degradation estimations

Date

2024-1

Author

Vardar, Batuhan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

407
views

1282
downloads

Cite This

Blind Super-Resolution methods aims to enhance the resolution of multiple different low resolution image counterparts of a single high resolution image with unknown degradation processes. With the advancements of various techniques in computer vision, many of the methods have been applied in Blind SR area and research has been mostly dominant with CNN based models. In order to take advantage of success of transformers in computer vision area, we integrate transformers and degradation representations for a Blind SR model. Furthermore, with the observation of lack of peak performance compared to Non-Blind SR models we separate the Blind SR problem. This sub-par performance in ideal settings is due to the inherent complexity in blind settings. To address this issue, we present a deblurring and denoising approach specifically for Blind SR which serves as a preprocessing method. Specifically our preprocessing method standardizes the low resolution images to a single degradation, which then can be enhanced with Non Blind SR methods to achieve Blind-SR performance. In addition to this approach we also use the same baseline architecture and propose an end to end Blind SR approach. We evaluate both our approaches on Set5, Set14, B100 and Urban100 datasets and compare the results with other state-of-the-art approaches using PSNR metric.

Subject Keywords

Blind super-resolution, Contrastive learning, Degradation learning, Vision transformers

URI

https://hdl.handle.net/11511/108517

Collections

Graduate School of Natural and Applied Sciences, Thesis