Robust Cost Volume Generation Method for Dense Stereo Matching in Endoscopic Scenarios

Area of Science:

• Computer Vision
• Medical Imaging
• Robotics

Background:

• Stereo matching in binocular endoscopy is challenging due to radiometric distortion from poor lighting.
• Traditional methods lack performance in difficult areas, while deep learning methods face generalizability and complexity issues.

Purpose of the Study:

• To develop a computationally efficient, non-deep learning stereo matching method for binocular endoscopic scenarios.
• To address radiometric distortion and improve matching reliability in homogenous regions.

Main Methods:

• Constructing an initial cost volume using radiometric invariant histogram of gradient angle and amplitude descriptors.
• Implementing a novel cross-scale propagation framework to enhance matching in homogenous regions without increasing computational cost.
• Combining the proposed method with the Local-Expansion optimization algorithm.

Main Results:

• The method achieves performance close to deep learning algorithms with significantly less computation (65x faster cost volume generation).
• Experimental results on the Middlebury Version 3 Benchmark show top performance among non-deep learning algorithms when combined with Local-Expansion.
• Accuracy on surgical endoscopic datasets and a custom binocular endoscope reaches the millimeter level, comparable to deep learning approaches.

Conclusions:

• The proposed non-deep learning method offers a computationally efficient and accurate solution for stereo matching in challenging endoscopic environments.
• This approach provides a viable alternative to deep learning methods, particularly where computational resources are limited or generalizability is a concern.