Beware of Aliases -- Signal Preservation is Crucial for Robust Image Restoration

TL;DR

This paper introduces BOA-Restormer, a transformer-based image restoration model that uses frequency domain operations to create alias-free paths, significantly improving robustness without sacrificing restoration quality.

Contribution

The paper proposes a novel alias-free architecture for image restoration transformers, enhancing robustness by integrating frequency domain processing in downsampling and upsampling.

Findings

Alias-free paths improve model robustness.

Frequency domain operations preserve high-frequency details.

BOA-Restormer outperforms traditional models in robustness tests.

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.