Deep Complex Spatio-Spectral Networks with Complex Visual Inputs

ICLR 2025 Conference Submission1276 Authors

Keywords: Deep Complex Newtworks, Complex-valued color transformation

TL;DR: A robust complex-valued approach in Spatio-spectral domain for multiple tasks on both real and complex data.

Abstract:

Complex-valued neural networks have attracted growing attention for their ability to handle complex-valued data with enhanced representational capacity. However, their potential in computer vision remains relatively untapped.
In this paper, we introduce Deep Complex Spatio-Spectral Network (DCSNet), a fully complex-valued token-based, end-to-end neural network designed for binary segmentation tasks. Additionally, our DCSNet encoder can be used for image classification in the complex domain. We also propose an invertible real-to-complex (R2C) transform, which generates two complex-valued input channels, complex intensity and complex hue, while producing complex-valued images with distinct real and imaginary components.
DCSNet operates in both spatial and spectral domains by leveraging complex-valued inputs and complex Fourier transform.
As a result, the complex-valued representation is maintained throughout DCSNet, and we avoid the information loss typically associated with Real$\leftrightarrow$Complex transformations. Extensive experiments show that DCSNet surpasses existing complex-valued methods across various tasks on both real and complex-valued data and achieves competitive performance compared to existing real-valued methods, establishing a robust framework for handling both data types effectively.

Primary Area: applications to 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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 1276