Answers are grounded in this paper and cite their sources.
Single image dehazing at high resolution must propagate scene-wide haze statistics efficiently while exploiting depth-related geometry without destabilizing training. Convolutional networks restore local detail but lack large-scale consistency, Transformer-style global modeling is computationally expensive, and geometry-aware methods can degrade when unreliable depth is injected too aggressively. We propose SSMDehazeNet, a hybrid visual state-space model with progressive depth modulation. A state-space branch propagates global haze statistics with linear complexity while a parallel local branch reconstructs structure at full resolution; the two are coupled inside a Parallel Hybrid State-Space Module so that both paths remain available at the same feature layer. Geometry is introduced progressively: a zero-initialized Spatial Depth Modulator gives it negligible influence early in training and lets its contribution be learned from data, and the depth-consistency objective is modulated by the current reconstruction residual. On the RESIDE SOTS benchmark, SSMDehazeNet achieves 42.48 dB/0.9968 (indoor) and 36.18 dB/0.9931 (outdoor) with favorable no-reference quality, and at 512 × 512 resolution it runs in 80.62 ms on an RTX 4090, substantially faster than heavy Transformer-based dehazing. Zero-shot evaluation on O-HAZE and NH-HAZE, a branch-level analysis, and a controlled geometry-corruption study further characterize its generalization and its reliance on the learned geometry cue.
No takes yet. Share an insight, caveat, or question.
Xie et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: