Revitalizing Channel-dimension Fourier Transform for Image Enhancement
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Image Enhancement, Fourier transform, Image Restoration
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Exploring the global representations of Fourier transform for image enhancement has become an alternative and made significant advancements. However, previous works only operate in the spatial dimensional, overlooking the potential of the channel dimension that inherently possesses discriminative features. In this work, we propose a fresh perspective, channel-dimension Fourier transform, for image enhancement. Our designs are simple yet effective and comprise three straightforward steps: applying the Fourier transform to the channel dimension to obtain channel-wise Fourier domain features, performing a channel-wise transformation on both its amplitude and phase components, and then reverting back to the spatial domain. Following the above rules, we offer three alternative implementation formats of the channel transform in different operational spaces, performing operations in 1) the global vector with higher orders; 2) the global vector with channel groups; and 3) the Fourier features derived from spatial-based Fourier transform. The above core designs, as general operators, can be seamlessly integrated with enhancement networks, achieving remarkable gains and building efficient models. Through extensive experiments on multiple image enhancement tasks, like low-light image enhancement, exposure correction, SDR2HDR translation, and underwater image enhancement, our designs exhibit consistent performance gains. The code will be publicly available.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6166
Loading