Light Field Image Super-Resolution Using Deformable Convolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Light field (LF) cameras capture multi-perspective views, offering angular information beneficial for super-resolution (SR).
• Incorporating this angular information for LF image SR is challenging due to inherent disparities between views.

Purpose of the Study:

• To propose a novel network, LF-DFnet, to effectively handle disparity issues in light field image super-resolution.
• To improve the incorporation and encoding of angular information for enhanced SR reconstruction across all views.

Main Methods:

• Developed a deformable convolution network (LF-DFnet) specifically for LF image SR.
• Introduced an angular deformable alignment module (ADAM) for feature-level alignment.
• Implemented a collect-and-distribute approach for bidirectional alignment between center and side-view features.

Main Results:

• LF-DFnet successfully incorporates and encodes angular information, improving SR reconstruction quality.
• Achieved state-of-the-art reconstruction accuracy with more faithful details compared to existing methods.
• Demonstrated superior robustness to varying disparity levels, a significant advancement in the field.

Conclusions:

• The proposed LF-DFnet effectively overcomes challenges posed by angular disparities in LF image SR.
• The method offers a robust and accurate solution for generating high-resolution light field images.
• LF-DFnet represents a significant step forward in light field super-resolution research.