Efficient per-pixel continuous representation and reconstruction of multi-view images

This paper proposes a novel representation of multi-view images. They consist of views of the same scene, captured from different camera positions. The proposal is optimal representation for compression, performance, resampling, and artifact mitigation. Two compression levels are proposed. The first reduces operation memory requirements with fast decoding, and the second for heavy compression on the disk. The main use case of the proposal is usage with 3D displays where different devices require different numbers and formats of the views. The required views can be quickly synthesized with artifacts and noise mitigation. Each pixel is represented by a continuous color function of all views using discrete cosine transform or circular harmonics. Higher quality is ensured in the areas in-focus on the 3D display. Artifacts in out-of-focus areas are not detectable. The proposal offers a better quality/bitrate ratio than competitors on noisy images, and still yields reasonable results on denoised images.