FACTNet: a frequency-domain-informed transformer for mmWave image computational super-resolution in the spatial frequency domain

Millimeter-wave (mmWave) imaging is a vital security screening technology, prized for its non-ionizing and penetrative capabilities. However, its efficacy is fundamentally limited by the physical principle of diffraction, which acts as a low-pass filter, irreversibly attenuating the high-frequency spatial details essential for identifying concealed objects. Prevailing deep learning-based computational super-resolution methods, operating predominantly in the spatial domain, struggle to counteract this frequency-specific information loss and thus often fail to restore fine textures and sharp edges. Inspired by this physical insight, we propose the frequency-aware cross-attention transformer network (FACTNet), an architecture designed to directly confront this challenge in the frequency domain. Its core innovation, the frequency-domain transformation module (FDTM), empowers the network to learn an adaptive filter in the Fourier domain, explicitly amplifying the high-frequency spectral components suppressed during image acquisition. Evaluations on our custom mmWave security dataset demonstrate that FACTNet achieves state-of-the-art performance, yielding reconstructions with superior visual fidelity, enhanced detail, and fewer artifacts. Our work establishes that aligning a network’s computational domain with the physical nature of image degradation provides a more effective and principled pathway to solving computational super-resolution reconstruction problems.