Conquering the Dark: LuxRA-Net for EBAPS Image Denoising from Full Moonlight to Overcast Starlight

Standard CMOS imagers provide reliable real-time imaging only above 1 lx, typical of daylight. Electron Bombarded Active Pixel Sensor (EBAPS), employing photoelectron bombardment gain, delivers clear real-time images down to > 10^-5 lx, covering full night illumination and extending operational range by five orders. Nevertheless, its distinct noise statistics remain underinvestigated, and targeted denoising frameworks are still scarce. To bridge this gap, we introduce the Multi-Illumination-Level Dataset (MILD), the first dedicated dataset for EBAPS imaging. Each scene in MILD contains seven synchronized images: one clean reference captured under normal lighting and six noisy counterparts spanning illumination levels from 10^-2 to 10^-5 lx. By maintaining fixed geometry and using aperture-controlled illumination, we ensure perfectly aligned clean-noisy pairs and reliable lux annotations. The dataset comprises 800 scenes and 4,800 high-resolution clean-noisy image pairs. Notably, as illumination decreases, EBAPS noise rises exponentially, rendering CMOS-oriented denoisers ineffective. To tackle this challenge, we propose LuxRA-Net, a lux-adaptive network integrating frequency-domain noise modeling with structural priors. Its core Frequency-Aware Noise Scaling (FANS) block extracts adaptive noise priors via spectral radial normalization. Together with texture enhancement and dynamic feature fusion, it captures the nonlinear relationship between noise and illumination intensity. Extensive experiments show LuxRA-Net outperforms state-of-the-art methods, achieving a 2.1 dB PSNR improvement and an 8% SSIM increase at 10^-5 lx, particularly excelling in edge preservation under photon-limited conditions. By introducing a new EBAPS benchmark and an illumination-adaptive denoiser, this work advances extreme low-light imaging and paves the way for next-generation night-vision applications.