Spatiotemporal frequency-domain filtering method for random row stripe noise in low-light short-wave infrared imaging

This is the authors' abstract. We don't add key points for this paper.

Indium gallium arsenide-based short-wave infrared (SWIR) imaging has long been recognized as an important approach for all-weather low-light imaging. However, after deep cooling suppresses detector dark current, electronics-induced random row stripe noise becomes dominant and appears as vertically rolling artifacts. Unlike static fixed-pattern noise, this amplitude-random, temporally dynamic row stripe noise cannot be effectively suppressed by conventional destriping methods. To address this issue, we analyze its generation mechanism and propose a spatiotemporal frequency-domain filtering method guided by the physical priors of random row stripe noise. The method reorganizes the degraded image sequence into independent row-temporal slices according to the row-wise coupling characteristic of the noise. Subsequently, it detects and isolates the moving targets using structure-tensor motion awareness to avoid motion artifacts during frequency-domain filtering. The motion-isolated slices are transformed to the frequency domain, where an inverted Gaussian notch filter is designed to suppress stripe noise based on its low-spatial-frequency prior. Finally, a 1D purification rule is applied to extract the stripe component and reconstruct the destriped image sequence. Experiments in four low-light SWIR scenes showed that the proposed method removes random row stripe noise more effectively than representative destriping methods while preserving image details. This study can facilitate the subsequent applications of low-light SWIR imaging technology.