MineSSM: Frequency-Decoupled State-Space Modeling for Real-Time Low-Light Enhancement in Dusty Underground Mines
TL;DR
MineSSM is a efficient design that routes each band to the operator it needs: a linear-complexity state-space model (Mamba) that homogenizes the global illumination and a lightweight convolutional branch that performs dust suppression on the high-frequency bands, so that no heavy operator ever runs at full resolution.
Abstract
Safety monitoring and visual surveillance in underground coal mines must cope with extreme low light and pervasive coal dust, yet the most accurate low-light image enhancement models rely on self-attention, whose quadratic cost precludes real-time use on the edge-computing platforms deployed underground, such as inspection robots and explosion-proof cameras. Our key observation is that these two degradations are separable by a wavelet decomposition: the low-frequency band carries the illumination to be corrected, whereas dust-induced noise is largely confined to the high-frequency bands. MineSSM turns this observation into an efficient design that routes each band to the operator it needs: a linear-complexity state-space model (Mamba) that homogenizes the global illumination and a lightweight convolutional branch that performs dust suppression on the high-frequency bands, so that no heavy operator ever runs at full resolution, and the two branches are recomposed under a Retinex constraint for faithful color restoration. A frequency-domain analysis of a real dusty mining frame confirms this split, with the smooth veil and lamp glow in the low-frequency band and the discrete dust speckles in the high-frequency bands. MineSSM surpasses the Transformer-based MEFormer on the MELOL mining dataset (26.58 dB PSNR/0.96 SSIM vs. 26.34 dB/0.91 SSIM) and attains the best PSNR and SSIM on the LOLv1 benchmark, while running at 27.8 FPS (0.036 s per 400×600 image) on a desktop GPU with linear resolution scaling. On an embedded NVIDIA Jetson Orin NX it sustains 8.0 FPS at 400×600, supporting near-real-time on-device enhancement at typical surveillance resolutions.