Blind Co-channel Interference Cancellation Using Fast Fourier Convolutions

Abstract

Addressing long-range dependencies in blind co-channel interference waveforms typically requires convolutional networks with large kernels or significant depth, which are resource-intensive. This paper presents a streamlined UNet architecture integrated with fast Fourier convolution blocks and a long short-term memory in the bottleneck, designed to efficiently capture these dependencies. By leveraging the Fourier domain for global feature processing, our architecture reduces the model's complexity without compromising performance. Compared to the leading benchmark model (a deep UNet), our approach yields a 26.5% improvement in mean square error, while reducing multiply-accumulate operations and the number of model parameters by 76.8% and 76.3% respectively, demonstrating a significant enhancement in both accuracy and efficiency for interference cancellation in constrained computational environments.