Color image guided depth image reconstruction based on a total variation network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Improving depth image quality is crucial for 3D reconstruction.
• Current color-guided depth reconstruction methods suffer from misalignment, causing texture artifacts and blurred discontinuities.
• High-resolution color images offer potential for enhanced depth recovery.

Purpose of the Study:

• To develop a novel method for color-guided depth reconstruction that overcomes existing limitations.
• To improve the accuracy of depth discontinuities and reduce artifacts in reconstructed depth images.
• To leverage high-resolution color images within a deep learning framework for superior depth estimation.

Main Methods:

• A total variation deep network was developed using deep learning principles.
• The network utilizes high-resolution color images to guide the depth image reconstruction process.
• The method focuses on aligning color details with depth discontinuities to mitigate artifacts.

Main Results:

• The recovered depth images show improved accuracy in edge contours compared to ground truth.
• The method effectively reduces texture copy artifacts and blurs at depth discontinuities.
• Significant retention of contour and positional information was observed even in low-resolution depth images.

Conclusions:

• The proposed deep learning approach effectively enhances depth image quality using high-resolution color guidance.
• The method demonstrates superior performance in preserving edge details and depth discontinuities.
• This technique offers a robust solution for accurate 3D information extraction from combined color and depth data.