Depth Super-Resolution via Joint Color-Guided Internal and External Regularizations.

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Depth information is crucial for numerous applications but is often captured at low resolution.
• Existing depth sensing technologies face limitations in producing high-resolution depth maps comparable to color images.

Purpose of the Study:

• To develop an effective depth super-resolution technique by integrating internal smoothness and external gradient consistency.
• To address the structural discrepancies between low-resolution depth maps and high-resolution guidance images.

Main Methods:

• A novel graph Laplacian regularizer is proposed to preserve the piecewise smooth nature of depth data.
• A graph gradient consistency constraint is introduced, aligning depth gradients with thresholded guidance image gradients.
• The approach unifies internal and external regularizations within a single optimization framework, solved using Alternating Direction Method of Multipliers (ADMM).

Main Results:

• The proposed method effectively leverages both depth data's inherent properties and guidance image information.
• Experimental results show superior performance compared to state-of-the-art methods in objective and subjective quality evaluations.
• The technique successfully remedies structural discrepancies between depth and guidance information.

Conclusions:

• The combined internal smoothness and external gradient consistency in a graph domain offers a robust solution for depth super-resolution.
• The developed method demonstrates significant improvements in reconstructing high-resolution depth maps from low-resolution inputs.
• This approach holds promise for advancing various real-world applications reliant on accurate depth data.