Stereo Image Restoration via Attention-Guided Correspondence Learning.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing stereo image restoration methods are limited by binocular symmetry, restricting them to horizontal parallax.
• Stereo images with unlimited parallax present significant challenges in real-world applications and remain underexplored.

Purpose of the Study:

• To develop a novel method for restoring high-quality stereo images with unlimited parallax.
• To address the challenges posed by large ranges and asymmetrical parallax in stereo image restoration.

Main Methods:

• Proposes an attention-guided correspondence learning method utilizing parallax and omnidirectional attention.
• Introduces the Selective Parallax Attention Module (SPAM) for adaptive cross-view feature interaction based on parallax.
• Develops the Non-local Omnidirectional Attention Module (NOAM) to capture global contextual correlations for asymmetrical parallax.

Main Results:

• The proposed Attention-guided Correspondence Learning Restoration Network (ACLRNet) effectively restores stereo images by learning feature correspondence.
• Demonstrated significant improvements in stereo image super-resolution, denoising, and artifact reduction.
• Achieved state-of-the-art performance across five benchmark datasets.

Conclusions:

• The proposed method successfully restores stereo images with unlimited parallax, outperforming existing approaches.
• The attention-guided strategy effectively learns feature correspondence, crucial for handling complex parallax.
• The method shows strong generalization capabilities across various stereo image restoration tasks.