Yuan-Kai Wang
  • Home
  • Research
    • Recent Interests >
      • Single Image Defogging
      • Night Image Enhancement
      • Camera Anomaly Detection
      • Accelerating Retinex by GPGPU Enabled Parallelization Algorithm
    • Publication
    • Research Grants
    • Group Members
  • Teaching
  • Activity
  • Industry
  • About Me

Single Image Defogging by  Multiscale Depth Fusion


Restoration of fog images is important for the de-weathering issue in computer vision. The problem is ill-posed and can be regularized within a Bayesian context by using a probabilistic fusion model. This paper presents a multiscale depth fusion (MDF) method for defog from a single image. A linear model representing the stochastic residual of nonlinear filtering is first proposed. Multiscale filtering results are probabilistically blended into a fused depth map based on the model. The fusion is formulated as an energy minimization problem that incorporates spatial Markov dependency. An inhomogeneous Laplacian-Markov random field for the multiscale fusion regularized with smoothing and edge-preserving constraints is developed. A nonconvex potential, adaptive truncated Laplacian, is devised to account for spatially variant characteristics such as edge and depth discontinuity. Defog is solved by an alternate optimization algorithm searching for solutions of depth map by minimizing the nonconvex potential in the random field. The MDF method is experimentally verified by real-world fog images including cluttered-depth scene that is challenging for defogging at finer details. Fog-free images are restored with improving contrast and vivid colors but without over-saturation. Quantitative assessment of image quality is applied to compare various defog methods. Experimental results demonstrate that the accurate estimation of depth map by the proposed edge-preserved multiscale fusion should recover high-quality images with sharp details.

Publication

  • Y.K. Wang and C.T. Fan, "Single Image Defogging by Multiscale Depth Fusion", IEEE Transactions on Image Processing, vol. 23, no. 11, pp. 4826 - 4837, 2014. [Pre-print] [IEEE Xplore]
  • Y.K. Wang, C.T. Fan, and C.W. Chang, "Accurate Depth Estimation for Image Defogging using Markov Random Field," in 4th International Conference on Graphics and Image Processing, Singapore, Mar. 2013, pp.87681Q. 

Experiments

Please "click" on an image to show details.

Synthetic Example

图片
Synthetic two-albedo image
图片
Synthetic transmission
图片
Estimated transmission by ILMRF
图片
Estimated transmission by He's method
图片
Estimated transmission by Fattal's method
图片
Synthetic foggy image
图片
Recovered image by MDF
图片
Recovered image by He's method
图片
Recovered image by Fattal's method

Halo effect of defogging incurred by block-based nonlinear filtering


Reduction of halo effect by depth fusion


Comparison of depth estimates for a cluttered-depth image.


Comparison of edge-preservation between guided filter and ILMRF


Restoration of a city image with haze


Restoration of a natural scene with clouds and mists


Restoration of a dense-fog image with cluttered depths segmented by trees


Supported Materials

Please "click" on an image to show details.

Comparison of depth estimates for a smooth-depth image


Restoration of a mountain image with clouds and mists.


Restoration of a red-brick house image with low visibility

Contact Me
Google Scholar
ORCID
LinkedIn
My Channel
(Chinese)
Intelligent
​System Lab
    
Icons made by Gregor Cresnar from www.flaticon.com is licensed by CC 3.0 BY
  • Home
  • Research
    • Recent Interests >
      • Single Image Defogging
      • Night Image Enhancement
      • Camera Anomaly Detection
      • Accelerating Retinex by GPGPU Enabled Parallelization Algorithm
    • Publication
    • Research Grants
    • Group Members
  • Teaching
  • Activity
  • Industry
  • About Me