Complex-Valued Deep Learning

Fully complex-valued neural architectures for signal and image processing with richer representational capacity.

Complex-valued deep learning extends conventional real-valued neural networks to operate entirely in the complex domain, where each neuron processes both the amplitude and phase of a signal. This yields richer representational capacity and is particularly well-suited for signals that inherently carry phase information — such as InSAR (Interferometric Synthetic Aperture Radar) imagery and RF signals. Our work introduces the first fully complex-valued architectures for visual perception and SAR-based building segmentation. Code

Fully Complex-Valued Deep Learning Model for Visual Perception (ICASSP 2023):

A fully complex-valued deep learning scheme for visual perception tasks. The network operates end-to-end in the complex domain, demonstrating improved representational capacity and performance over real-valued counterparts.

FC2MFN: Fully Complex-Valued Fully Convolutional Multi-Feature Fusion Network for InSAR Segmentation (SSCI 2022):

FC2MFN is a fully complex-valued fully convolutional network with multi-feature fusion for building segmentation from InSAR imagery, leveraging complex-valued operations throughout to preserve phase coherence.