Sparse Depth-Guided Image Enhancement Using Incremental GP with Informative Point Selection.

Area of Science:

• Computer Vision
• Image Processing
• Robotics

Background:

• Single image dehazing traditionally relies on multi-channel information.
• Robotics platforms often provide sparse range measurements, a valuable but underutilized data source for image processing.

Purpose of the Study:

• To develop an efficient and effective online image dehazing method leveraging sparse depth priors.
• To enable robust dehazing for both color and grayscale images using readily available, albeit sparse, depth data.

Main Methods:

• An incremental Gaussian Process (iGP) was employed for efficient, incremental depth map estimation.
• Depth prior selection was guided by information-theoretic metrics, specifically mutual information (MI).
• Haze-free images were reconstructed using the atmospheric scattering model with incrementally estimated depth.

Main Results:

• The proposed method successfully dehazed color and grayscale images in various scenarios, including synthetic fog and real-world indoor, outdoor, and underwater conditions.
• The algorithm demonstrated effectiveness even with highly sparse depth priors.
• The iGP approach facilitated efficient online processing and reconstruction.

Conclusions:

• Exploiting sparse depth priors with iGP offers an efficient and effective solution for online image dehazing.
• The method's applicability to both color and grayscale images broadens its utility in diverse applications, particularly in robotics.
• Information-theoretic selection of depth priors enhances the performance of the incremental dehazing process.