PaperlayerView exact revision on arXiv

Source-linked AI summary

Night Time Haze and Glow Removal using Deep Dilated Convolutional Network

Shiba Kuanar, K. R. Rao, Dwarikanath Mahapatra, Monalisa Bilas

arXiv:1902.00855v1cs.CV

TL;DR

Nighttime haze removal is complicated by colored, non-uniform illumination and glow from visible light sources. The paper proposes a recurrent DeGlow-DeHaze network that progressively removes glow and haze, outperforming state-of-the-art methods in image quality and computation time.

  • Problem

    Nighttime haze removal must address haze alongside glow from visible light sources with varying colors, shapes, and non-uniform illumination.

    Source: §3.1 · p. 4

  • Method

    A recurrent dilated DeGlow network detects and removes glow before a separate DeHaze model suppresses haze and restores visibility.

    Source: §3.1 · p. 5, §3.3 · p. 6, §3.5 · p. 8

  • Results

    The model outperforms state-of-the-art methods on PSNR, SSIM, and computation time across synthetic and real nighttime images.

    Source: §4.2 · p. 11

  • Takeaways & Limitations

    The approach corrects glow and color distortion while enhancing visibility with subtle artifacts in real and synthetic nighttime images.

    Source: §4.1 · p. 10, §4.2 · p. 11

    The method can generate color distortions and fail when objects are occluded or surrounded by large objects, causing glows to merge or blow out.

    Source: §4.2 · p. 11

Abstract

from arXiv · show

In this paper, we address the single image haze removal problem in a nighttime scene. The night haze removal is a severely ill-posed problem especially due to the presence of various visible light sources with varying colors and non-uniform illumination. These light sources are of different shapes and introduce noticeable glow in night scenes. To address these effects we introduce a deep learning based DeGlow-DeHaze iterative architecture which accounts for varying color illumination and glows. First, our convolution neural network (CNN) based DeGlow model is able to remove the glow effect significantly and on top of it a separate DeHaze network is included to remove the haze effect. For our recurrent network training, the hazy images and the corresponding transmission maps are synthesized from the NYU depth datasets and consequently restored a high-quality haze-free image. The experimental results demonstrate that our hybrid CNN model outperforms other state-of-the-art methods in terms of computation speed and image quality. We also show the effectiveness of our model on a number of real images and compare our results with the existing night haze heuristic models.

3.1. Equations for Glow Detection and Removal · AUTHOR INFORMATION:

The paper models nighttime glow and haze jointly, progressively removing glow before applying DeHaze restoration. Its recurrent multi-path dilated framework estimates scene components and transmission while improving visibility, color balance, and computational efficiency, with failure cases under occlusion.

3.1. Equations for Glow Detection and Removal

  • The DeGlow-DeHaze framework first estimates and subtracts glow from the input, then removes residual haze and estimates atmospheric light from the transmission map.It forms J(x) = I(x) − G(x), applies a separate DeHaze model, and selects atmospheric light from the top 0.1% darkest transmission-map pixels.Source: §3.1 · p. 5
  • Three dilated convolution paths with DF values 1, 2, and 3 provide receptive fields of 7×7, 13×13, and 17×17 while preserving resolution.The recurrent structure expands contextual coverage across iterations, and each path contains three 3×3 convolutions.Source: §3.1 · p. 5

3.3. CNN Based Recurrent DeGlow Network

  • The recurrent DeGlow network jointly estimates J(x), S(x), and G(x) through MAP-regularized inverse recovery from the observed glow image I(x).Its recursive predictions use dilated features and feedback among glow, scene, and haze-related components.Source: §3.3 · p. 6

3.4. DeGlow Network Training

  • The DeGlow training objective jointly learns glow, scene, and haze estimates over recurrent iterations using stochastic-gradient optimization with regularization.Training uses batch size 128, momentum 0.9, weight decay 0.001, and converges after approximately 1200 iterations in about eighteen hours on an NVIDIA K40 GPU.Source: §3.4 · pp. 6–7

3.5. Dehazing with Multi-Scale CNN Network

  • The one-recurrence DeHaze network is trained on synthesized haze images and transmission maps to suppress bright glow, increase contrast, and improve visibility.Training data include 45,000 synthesized hazy images generated from 5,000 NYU Depth images with varied scattering and glow parameters.Source: §3.5 · p. 8

4.1. Quantitative Comparison on Real Night Images

  • Across real and synthetic images, the proposed method achieved the highest SSIM among compared CNN models and outperformed other CNN approaches overall.Visual comparisons report reduced glow, normalized color distortion, enhanced visibility, and fewer artifacts than the compared methods.Source: Figure 7 · p. 9, §4.1 · p. 10

4.2. Failure Case

  • The method can produce color distortions and fail to eliminate merged or blown-out glows when objects are occluded or surrounded by large scene structures.The reported failure motivates improved color-constancy methods and datasets for occluded nighttime scenes.Source: Figure 10 · p. 11, §4.2 · p. 11

Related papers

Loading 1902.00855v1…