DIC-DDA: Learned Asymmetric Distributed Image Compression via Dual Domain Alignment.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Multi-view and stereo image compression are vital for 3D applications.
• Exploiting view correlations improves compression rates, but joint encoding is computationally intensive.
• Learned Distributed Image Coding (DIC) shifts complexity to the decoder using correlated side images (SI).

Purpose of the Study:

• To develop an asymmetric DIC method using one losslessly compressed side image (SI).
• To address the key challenge of accurate alignment between low-quality target images and high-quality SIs in learned asymmetric DIC.
• To improve alignment accuracy over traditional patch-level methods that rely on degraded SI.

Main Methods:

• Proposed a dual domain alignment strategy for pixel-wise offset estimation.
• Implemented degraded domain alignment by matching features of degraded SI and target images.
• Introduced fused domain alignment using the fusion of degraded target and aligned SI features to estimate fine-scale offsets.
• Developed a selective enhancement module to correct mismatched regions.

Main Results:

• The proposed dual domain alignment strategy significantly improves alignment accuracy.
• The method outperforms existing approaches, achieving 16% average BD-rate reduction on the KITTI Stereo dataset.
• Extensive experiments on three datasets validate the method's superiority.

Conclusions:

• The novel dual domain alignment strategy effectively addresses the alignment problem in learned asymmetric DIC.
• The proposed method offers superior compression performance for multi-view images.
• Future work includes releasing the code upon acceptance.