Enhancing Visibility in Nighttime Haze Images Using Guided APSF and Gradient Adaptive Convolution

TL;DR

This work tackles nighttime haze by combining glow suppression and low-light enhancement. It introduces an APSF-guided glow rendering pipeline and a Light Source Aware Network to model glow sources, paired with Gradient Adaptive Convolution to preserve edges and textures, and an attention-based module for boosting dim regions. The approach is trained in a CycleGAN framework with specialized losses (L_ls, L_g, L_k) and leverages semi-supervised learning through synthetic glow rendering, achieving a PSNR of $30.383$ dB on GTA5 nighttime haze and outperforming prior methods. The results demonstrate improved visibility, reduced glow, and preserved structural details, with strong potential for real-world nighttime vision tasks such as surveillance and autonomous driving.

Abstract

Visibility in hazy nighttime scenes is frequently reduced by multiple factors, including low light, intense glow, light scattering, and the presence of multicolored light sources. Existing nighttime dehazing methods often struggle with handling glow or low-light conditions, resulting in either excessively dark visuals or unsuppressed glow outputs. In this paper, we enhance the visibility from a single nighttime haze image by suppressing glow and enhancing low-light regions. To handle glow effects, our framework learns from the rendered glow pairs. Specifically, a light source aware network is proposed to detect light sources of night images, followed by the APSF (Atmospheric Point Spread Function)-guided glow rendering. Our framework is then trained on the rendered images, resulting in glow suppression. Moreover, we utilize gradient-adaptive convolution, to capture edges and textures in hazy scenes. By leveraging extracted edges and textures, we enhance the contrast of the scene without losing important structural details. To boost low-light intensity, our network learns an attention map, then adjusted by gamma correction. This attention has high values on low-light regions and low values on haze and glow regions. Extensive evaluation on real nighttime haze images, demonstrates the effectiveness of our method. Our experiments demonstrate that our method achieves a PSNR of 30.38dB, outperforming state-of-the-art methods by 13% on GTA5 nighttime haze dataset. Our data and code is available at https://github.com/jinyeying/nighttime_dehaze.

Figures (14)

• Figure 1: (1) Our deglowing framework $G_c$ have two inputs: one to learn from real haze images $I_h$ and the other to learn from real clean reference images $I_c$. For input haze images $I_h$, $G_c$ output clean images $O_c$. For input clean images $I_c$, $G_c$ output clean images $G_c(I_c)$. (2) APSF guide glow generator $G_h$ to generate glow $O_h$ on reference images $I_c$. (3) the upper left is the gradient adaptive convolution, from the gradient convolution (the blue window), we obtain edges; from the adaptive bilateral kernel (the red), we enhance texture details. (4) the upper right is attention-guided enhancement module.
• Figure 2: We show \ref{['algorithm_ls']}, light source map detection: (1) We first generate an initial light source mask $\hat{M}$ based on intensity, (2) then refine the mask using alpha matting levin2007closed to obtain light source soft matting ${M}$. (3) By multiplying the light source map ${M}$ with the night clean image $I_c$, we obtain the light source map $L_s$. After obtaining the light source, we show \ref{['algorithm_apsf']}, APSF-guided nighttime glow rendering: (1) Next, we perform APSF 2D convolution on the light source map to render glow $G$. (2) Finally, by combining the night clean and glow image, we obtain the rendered glow image $I_g$. More results are shown in \ref{['fig:apsf']}.
• Figure 3: We show the light source maps $L_s$ of night haze.
• Figure 4: We show with APSF, we can render glow $I_g$ (bottom) on night clean $I_c$ (top), with the help of light source maps.
• Figure 5: We show using gradient adaptive convolution, can obtain edges (middle) and textures (below) in haze images.