Researchers have introduced SSMDehazeNet, a hybrid visual state-space model designed for single image dehazing that couples linear-complexity global modeling with progressive depth modulation. The framework addresses persistent trade-offs in high-resolution image restoration by simultaneously propagating scene-wide haze distribution statistics and reconstructing fine local structural details without destabilizing network training.
High-resolution dehazing has historically presented significant computational and structural challenges for automated vision systems. Existing convolutional neural networks effectively restore local texture and edges but fail to maintain large-scale consistency across an entire image. Conversely, Transformer-based architectures offer global modeling capacity but remain computationally expensive, while conventional geometry-aware approaches frequently suffer visual degradation when unreliable depth information is introduced too aggressively into the restoration pipeline.
To overcome these limitations, SSMDehazeNet organizes feature extraction into a Parallel Hybrid State-Space Module that maintains both processing paths at the same feature layer. Inside this module, a visual state-space branch propagates global haze statistics across the entire scene with linear complexity. Operating alongside it, a parallel local branch reconstructs spatial structure at full resolution, allowing the network to retain fine visual elements while monitoring widespread atmospheric interference.
The system incorporates geometric cues gradually rather than imposing rigid three-dimensional priors at the outset of optimization. A zero-initialized Spatial Depth Modulator ensures that depth information exerts negligible influence during initial training phases, enabling the model to learn geometry contributions progressively from training data. Furthermore, the researchers designed a depth-consistency objective that is dynamically modulated by the current reconstruction residual.
When tested on the RESIDE SOTS benchmark, SSMDehazeNet achieved restoration scores of 42.48 dB and 0.9968 on indoor images, as well as 36.18 dB and 0.9931 on outdoor images, alongside favorable no-reference quality metrics. Processed at 512 by 512 resolution on an Nvidia RTX 4090 graphics processing unit, the architecture recorded an inference runtime of 80.62 milliseconds, performing substantially faster than heavy Transformer-based models. The authors also characterized model generalization and dependence on geometric cues through zero-shot evaluations on O-HAZE and NH-HAZE benchmarks, branch-level analyses, and controlled geometry-corruption tests.
Subject of Research: Single image dehazing using hybrid visual state-space models
Article Title: A hybrid visual state-space model with progressive depth modulation for single image dehazing
Article References: Xie, Z., & Zhao, H. (2026). A hybrid visual state-space model with progressive depth modulation for single image dehazing. Scientific Reports. https://doi.org/10.1038/s41598-026-78372-x
Image Credits: AI Generated
DOI: 10.1038/s41598-026-78372-x
Keywords: single image dehazing, SSMDehazeNet, visual state-space model, progressive depth modulation, Parallel Hybrid State-Space Module, Spatial Depth Modulator, depth-consistency objective, RESIDE SOTS benchmark, O-HAZE dataset, NH-HAZE dataset, image restoration, linear complexity, convolutional networks, vision Transformers, reconstruction residual, high-resolution dehazing
Cite Scienmag News
Cassandra Pierce. (November 22, 2026). Hybrid State-Space Model Clears Hazy Images Using Progressive Depth Modulation. Scienmag. https://scienmag.com/hybrid-state-space-model-clears-hazy-images-using-progressive-depth-modulation/
Cassandra Pierce. "Hybrid State-Space Model Clears Hazy Images Using Progressive Depth Modulation." Scienmag, 22 November 2026, https://scienmag.com/hybrid-state-space-model-clears-hazy-images-using-progressive-depth-modulation/. Accessed 22 November 2026.
Cassandra Pierce. "Hybrid State-Space Model Clears Hazy Images Using Progressive Depth Modulation." Scienmag. November 22, 2026. https://scienmag.com/hybrid-state-space-model-clears-hazy-images-using-progressive-depth-modulation/








