Depth Estimation for Light-Field Images Using Stereo Matching and Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Light-field (LF) imaging captures scene geometry and appearance.
• Accurate depth estimation from LF images is crucial for various applications.
• Existing methods often struggle with precision and computational efficiency.

Purpose of the Study:

• To develop a novel, accurate, and efficient depth-estimation method for light-field images.
• To leverage multi-stereo matching and deep learning for enhanced disparity map generation.
• To outperform current state-of-the-art machine learning-based depth estimation techniques.

Main Methods:

• A two-stage approach combining block-based stereo matching with deep learning (DL).
• Initial disparity estimation using a novel multi-stereo matching algorithm across sub-aperture images (SAIs).
• Refinement of disparity maps via a pixel-wise DL-based residual error prediction using a novel neural network architecture.

Main Results:

• The proposed method significantly improves depth estimation accuracy.
• Achieved average improvements of 15.65% in RMSE, 43.62% in MAE, and 5.03% in SSIM.
• Demonstrated superior performance compared to existing machine learning-based state-of-the-art methods.

Conclusions:

• The proposed hybrid approach effectively enhances depth estimation for light-field images.
• The novel neural network architecture and stereo matching techniques contribute to improved accuracy.
• This method offers a promising advancement in light-field depth estimation technology.