Depth Estimation for Integral Imaging Microscopy Using a 3D-2D CNN with a Weighted Median Filter.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Depth map estimation is crucial for 3D scene understanding and computer vision applications.
• Existing methods often struggle with accuracy, especially at object boundaries.
• Deep learning approaches offer potential for improved depth estimation performance.

Purpose of the Study:

• To propose a novel and robust depth map framework using convolutional neural networks (CNNs).
• To enhance depth estimation accuracy by leveraging multi-direction epipolar plane images (EPIs).
• To address noise and improve boundary information in depth maps.

Main Methods:

• Utilized a combination of 3D and 2D CNNs for feature extraction from multi-direction EPIs.
• Employed adapted 3D convolutional blocks for varying epipolar image disparities.
• Integrated 2D CNNs to minimize data loss and a fully convolutional approach for scalability.
• Applied weighted median filtering (WMF) to refine boundary information and reduce noise.

Main Results:

• The proposed deep learning framework demonstrated superior depth estimation accuracy compared to existing architectures.
• The multi-stream network effectively merged features to restore depth information.
• Weighted median filtering successfully improved accuracy by addressing edge noise.

Conclusions:

• The developed CNN-based framework provides a robust solution for accurate depth map generation.
• The integration of multi-direction EPIs and advanced filtering techniques leads to significant performance gains.
• This approach offers a scalable and accurate method for depth estimation in various applications.