Quad-directional frequency-aware state space modeling for low-light image enhancement

Low-Light Image Enhancement (LLIE) is a fundamental inverse problem in computational photography characterised by spatially non-uniform underexposure, amplified photon shot noise, and inter-channel colour imbalance. State-of-the-art (SOTA) Transformer based methods achieve compelling global context modelling but impose quadratic self-attention complexity, precluding efficient deployment at high resolutions. We present QDFA-SSM, a novel end-to-end framework built entirely on selective State Space Models (SSMs). Three principled architectural components constitute our design. First, a Quad-Directional Visual Mamba Encoder (QD-VME) scans 2D feature maps along four spatial orientations horizontal forward/backward and vertical forward/backward recovering isotropic long-range illumination context at strictly $$\mathscr {O}(N)$$ complexity via hardware efficient parallel scanning. Second, a Frequency-Aware SSM Decoder (FA-SSMD) decomposes decoder features into low- and high-frequency spectral bands via 2D FFT, applying independent SSM chains to each band to separately optimise global luminance consistency and local texture fidelity, preventing cross-band interference. Third, a Spatially-Adaptive Illumination Modulator (SAIM) estimates pixel-wise enhancement weight maps conditioned on local Signal-to-Noise Ratio (SNR) maps, concentrating correction capacity on severely underexposed regions without requiring additional supervision. Evaluated on the LOL-v1 benchmark, QDFA-SSM attains PSNR = 25.15 dB and SSIM = 0.863, which improves upon the strongest prior selective-SSM baselines MMamba-LLIE and LatentDark by $$+0.67$$ dB and $$+0.42$$ dB respectively, and upon the Retinex-guided Transformer baselines represented by Retinexformer and LLFormer. These gains are obtained with 1.340 M parameters and 32.105 GFLOPs at 41.53 FPS on an NVIDIA Tesla T4 GPU. It may be noted that the proposed model is not the smallest or the fastest among the compared methods; rather, it is reported here as offering the most favourable balance between restoration accuracy and computational cost.