Abstract
Single-image dehazing often fails in scenarios where haze is distributed unevenly across the depth of field. This paper presents a novel approach to address "disproportionate haze" by integrating dual-patch transmission map estimations. By combining large-patch stability for foregrounds with small-patch sensitivity for backgrounds through Laplacian pyramid fusion and saliency masking, the proposed method achieves superior restoration quality without the artifacts typical of single-scale approaches.
Introduction
The Dark Channel Prior (DCP) is a cornerstone of modern image dehazing, yet it faces a fundamental limitation regarding patch size. In scenes with disproportionate haze—where the background is heavily obscured while the foreground remains relatively clear—a fixed patch size cannot optimize for both regions. Large patches (e.g., 15x15) often underestimate background haze, while small patches (e.g., 3x3) introduce halo artifacts in the foreground. This research proposes a hybrid framework to resolve this trade-off.
Methodology
The proposed system consists of two primary modules: Haze Thickness Estimation (HTE) and Haze Formation Removal (HFR).
Haze Thickness Estimation (HTE)
The HTE module generates two distinct transmission maps. A 15x15 patch size is employed to ensure foreground stability, preventing the over-processing of dark objects. Simultaneously, a 3x3 patch size is used to capture the dense haze in distant background regions.
Saliency-Based Laplacian Fusion
To merge these maps, a binary saliency mask identifies foreground objects. Rather than a simple linear blend, the authors utilize Laplacian pyramid fusion. This multi-scale decomposition allows for the integration of the two maps across different frequency bands, effectively eliminating edge artifacts and visible seams.
Image Restoration (HFR)
The final HFR module applies the atmospheric scattering model using the refined hybrid transmission map to recover the latent haze-free image.
Experimental Results
The method was evaluated on real-world sequences ("Animal," "People," and "Scene") using both subjective and objective metrics.
Performance Metrics
The proposed method achieved a Mean Opinion Score (MOS) of 3.81 from 100 human evaluators, outperforming standard DCP (2.17–3.13) and other state-of-the-art methods like Gao et al. (3.76). Quantitatively, it recorded the lowest (best) BRISQUE score of 40.47, indicating higher spatial quality and naturalness compared to competitors which ranged from 40.97 to 44.90.
Discussion and Conclusion
The hybrid patch-size approach effectively bridges the gap between background clarity and foreground integrity. While the method shows significant improvements, its performance is contingent on the accuracy of the saliency mask and involves higher computational complexity due to the multi-scale fusion process. Future work may focus on optimizing these components for real-time applications and improving robustness in sky-dominated regions.
Summary updated