aiqbench analysis
This research addresses the trade-off between high PSNR scores and model robustness in image restoration by introducing alias-free signal paths within transformer architectures. By executing downsampling and upsampling partly in the frequency domain, the proposed BOA-Restormer maintains critical high-frequency information while significantly improving resistance to adversarial attacks without sacrificing performance.
Beware of Aliases -- Signal Preservation is Crucial for Robust Image Restoration
arXiv:2406.07435v2 Announce Type: replace-cross Abstract: Image restoration networks are usually comprised of an encoder and a decoder, responsible for aggregating image content from noisy, distorted data and to restore clean, undistorted images, respectively. Data aggregation as well as high-resolution image generation both usually come at the risk of involving aliases, i.e.~standard architectures put their ability to reconstruct the model input in jeopardy to reach high PSNR values on validation data. The price to be paid is low model robustness. In this work, we show that simply providing alias-free paths in state-of-the-art reconstruction transformers supports improved model robustness at low costs on the restoration performance. We do so by proposing BOA-Restormer, a transformer-based image restoration model that executes downsampling and upsampling operations partly in the frequency domain to ensure alias-free paths along the entire model while potentially preserving all relevant high-frequency information.
Abstract:Image restoration networks are usually comprised of an encoder and a decoder, responsible for aggregating image content from noisy, distorted data and to restore clean, undistorted images, respectively. Data aggregation as well as high-resolution image generation both usually come at the risk of involving aliases, i.e.~standard architectures put their ability to reconstruct the model input in jeopardy to reach high PSNR values on validation data. The price to be paid is low model robustness. In this work, we show that simply providing alias-free paths in state-of-the-art reconstruction transformers supports improved model robustness at low costs on the restoration performance. We do so by proposing BOA-Restormer, a transformer-based image restoration model that executes downsampling and upsampling operations partly in the frequency domain to ensure alias-free paths along the entire model while potentially preserving all relevant high-frequency information.
Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Image and Video Processing (eess.IV)
arXiv:2406.07435 [cs.CV]
(or arXiv:2406.07435v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2406.07435 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 239 words
Related coverage
Think Before You Restore: Risk-Aware Manchu Manuscript Restoration with Stroke-Guided Attention
arXiv · Sep 30, 2026
Inverse Manipulation through Symbolic Planning and Residual Operator Learning
arXiv · Jun 5, 2026
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
arXiv · Sep 1, 2026
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
arXiv · Jun 10, 2026