Beware of Aliases -- Signal Preservation is Crucial for Robust Image Restoration
This work addresses robustness issues in image restoration for applications requiring reliable reconstruction, though it is incremental as it builds on existing transformer methods.
The paper tackles the problem of low robustness in image restoration networks caused by aliasing in standard architectures, and shows that providing alias-free paths in transformers improves model robustness with minimal impact on restoration performance.
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.