SPECTRA-Net: Spectral Recursive Amplification and Phase Calibration for Low-Light Image Enhancement

Low-light image enhancement (LLIE) remains a fundamental yet challenging problem in computer vision since existing methods often fail to simultaneously achieve computational efficiency, structural fidelity, and chromatic consistency, particularly in severely dark regions where latent scene structures are difficult to recover. In this paper, we propose SPECTRA-Net (Spectral-domain Recursive Amplification Network), a novel hybrid frequency spatial framework that follows a Boost then Calibrate paradigm to reveal hidden details in deep shadows while maintaining global structural coherence and natural color reproduction. Recognizing that critical structural information is frequently lost in the spatial domain under extreme low-light conditions, SPECTRA-Net explicitly operates in the Fourier domain and decouples luminance-related amplitude from structure-related phase information. To extract cleaner and more informative spectral representations from deep shadows, the amplitude is progressively and recursively boosted, deliberately driving the intermediate features into an over-exposed state to fully manifest latent details buried near the noise floor. Building upon this intentional over-exposure, we introduce a coupled amplitude-phase modulation mechanism in which the phase is actively calibrated by a linear-complexity VMamba2-based processor that captures spectral structural dependencies through selective state-space duality (SSD), thereby mitigating phase-amplitude mismatch and suppressing ringing artifacts. Subsequently, a Brightness-aware Attention Calibration module re-normalizes illumination into a natural perceptual range without discarding recovered details. SPECTRA-Net achieves state-of-the-art performance across multiple LLIE benchmarks, demonstrates superior visual clarity, robust structural preservation, and stable performance across both paired benchmarks (LOL, LOL-v2) and unpaired real-world datasets (DICM, LIME).