Real-World Scene Recovery for Scattering-Degraded Images Using Spatial and Frequency Priors

Technical Summary: Real-World Scene Recovery for Scattering-Degraded Images Using Spatial and Frequency Priors

Problem Statement
Scene recovery from images degraded by scattering effects—such as haze, sandstorms, underwater conditions, and remote sensing environments—remains a fundamental challenge in computer vision. These degradations introduce additive veiling components and reduce direct transmission, leading to reduced contrast, color distortion, and loss of structural details. This significantly impairs visual perception and the performance of downstream tasks like object detection and segmentation. Existing solutions generally fall into two categories: prior-based methods and learning-based methods. Prior-based approaches often rely on a single prior designed for specific scenarios (e.g., daytime haze), failing to generalize across diverse, heterogeneous scattering degradations. Conversely, learning-based methods, while powerful, typically require large-scale synthetic training data, suffer from domain gaps when applied to real-world scenarios, and demand substantial computational resources.

Methodology: Spatial and Frequency Priors (SFP)
The authors propose a novel framework, Spatial and Frequency Priors (SFP), designed to recover scenes from real-world scattering-degraded images without relying on large-scale training data. The framework operates through three main stages:

1. Spatial-Domain Prior (SDP):

   • Observation: The authors observe that the inverse of a scattering-degraded image ($1 - I(x)$) reveals a projection along its spectral direction that correlates with the underlying scene transmission.
   • Mechanism: The spectral direction $S(x)$ is modeled as the gradients of the three color channels. The method projects the inverted degraded image onto this spectral direction to estimate the scene transmission map $t(x)$.
   • Restoration: Once $t(x)$ and the global atmospheric light $A$ are estimated, the scene radiance $J(x)$ is recovered by inverting the Atmospheric Scattering Model (ASM).

2. Frequency-Domain Priors (FDP):

   • Observation 1 (DC Components): In clear images, the mean intensity of the Direct Current (DC) components across channels approximates the mean DC value of the corresponding degraded image.
   • Observation 2 (Low Radial Frequencies): In clear images, low radial frequencies (within a narrow band, $\rho(x) < 0.001$) contribute only a small proportion (approx. 1%) of the overall spectrum. Higher proportions indicate severe degradation.
   • Mechanism: Based on these observations, an adaptive mask $M_C(x)$ is constructed to perform frequency-domain enhancement. The mask parameters are adaptively determined to align the DC components and restore the low-frequency proportion to the expected 1% baseline, effectively compensating for scattering-induced attenuation.

3. Weighted Fusion Strategy:

   • The final recovered image is obtained by fusing the spatial-domain restored image, the frequency-domain enhanced image, and the original input.
   • The fusion is performed in the Lab color space. For the $a$ and $b$ channels, weights are assigned based on color neutrality (mean values near zero). For the $L$ (luminance) channel, a discrete wavelet transform (DWT) is used, retaining low-frequency components from the spatial result and fusing high-frequency components using a maximum-value strategy.
   • Post-processing (gamma correction and HDR compression) is applied to improve visual consistency.

Key Contributions

• Novel Framework: The introduction of SFP, which jointly exploits complementary priors in both spatial and frequency domains, moving beyond the limitations of single-domain or single-prior approaches.
• New Priors:
  • A spatial prior that leverages spectral properties to project the inverted degraded image onto the spectral direction, approximating scene transmission.
  • Two frequency priors utilizing the statistical characteristics of DC components and low radial frequencies to construct an adaptive mask for degradation-aware enhancement.
• Generalization: The method achieves superior performance across diverse real-world scenarios (haze, sandstorm, underwater, remote sensing) without requiring large-scale synthetic training datasets.
• Plug-and-Play Capability: The frequency-domain prior can be integrated with existing spatial-domain priors to improve their performance, demonstrating flexibility.

Experimental Results
Extensive experiments were conducted on diverse real-world datasets, including FTD, DHQ, HSTS, Fattal (haze); DID, DAWN (sandstorm); UIEB, U45, UCCS, UHTS (underwater); and UAV, RRSHID, RRSD (remote sensing). Additionally, the method was tested on mixed degradation datasets (RUSH) and complex nighttime hazy scenes (NHRW, RealNightHaze).

• Quantitative Performance: SFP consistently outperformed over 50 state-of-the-art methods (both prior-based and learning-based) across multiple metrics (NIMA, FADE, NIQE, UCIQE). For instance, on real-world hazy datasets, it improved average NIMA scores by approximately 0.66% and reduced FADE by 17.64%. On underwater datasets, it improved UCIQE scores by 6.70%.
• Visual Quality: Visual comparisons demonstrated that SFP effectively removes haze and color bias while preserving fine structural details and maintaining natural color appearance, avoiding the halo artifacts or over-enhancement common in other methods.
• Downstream Tasks: The framework significantly improved performance in high-level vision tasks. When used as a preprocessing step for object detection (YOLOv8), SFP achieved the highest average precision across categories (Person, Bicycle, Car, Motorbike, Bus) compared to other dehazing methods. It also improved results in keypoint detection and image segmentation.

Significance and Claims
The paper claims that SFP addresses the critical limitations of existing methods: the lack of adaptability in single-domain priors and the poor generalization of data-driven methods to real-world scenarios. By jointly modeling degradation characteristics in both spatial and frequency domains, SFP provides a robust, parameter-free solution that does not rely on heavy computational resources or synthetic data. The authors emphasize that the method is particularly effective in complex, coupled degradation scenarios, such as nighttime haze, and serves as a reliable preprocessing step to enhance the reliability of vision-based systems in real-world applications.