Video Compression GAN Upsampling Bandwidth Quality

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current systems for generating real-time volumetric video from multiple cameras require high computational power and suffer from high data transfer rates, necessitating significant compression, which often results in low perceptual quality due to hardware limitations and the need for high bandwidth, compromising immersive user experiences.

Innovation Solution

The implementation of a decompression Generative Adversarial Network (GAN) is used to upsample and improve the perceptual quality of compressed video streams, allowing for region-dependent downsampling and encoding, reducing the bitrate, and enhancing texture and sharpness, while maintaining high resolution for 3D model generation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of energy

If high compression ratios are applied to reduce data transfer rates, then bandwidth requirements are reduced, but perceptual quality deteriorates

Engineering Contradiction:
Improvebandwidth consumptionVSAvoidperceptual quality
Core Design Contradiction:
Loss of energyVSManufacturing precision

Solution Approach 1:

The patent changes the parameter space by training GANs on specific video compression scenarios to learn optimal reconstruction parameters. The generator network adjusts compression parameters dynamically to maintain perceptual quality while achieving high compression ratios, directly addressing the contradiction between bandwidth reduction and quality preservation.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces traditional mechanical/image processing compression systems with a learning-based GAN system. Instead of using conventional compression algorithms that prioritize file size, the GAN-based system uses neural network mechanisms to prioritize perceptual quality metrics, fundamentally substituting the compression approach to resolve the quality-bandwidth tradeoff.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Loss of energy

If region-dependent downsampling is applied to reduce bitrate, then bandwidth requirements are reduced, but uniform quality across all regions is lost

Engineering Contradiction:
Improvebandwidth consumptionVSAvoidquality uniformity
Core Design Contradiction:
Loss of energyVSManufacturing precision

Solution Approach 1:

The patent explicitly applies local quality principles by dividing the video frame into different regions and applying distinct downsampling strategies to each region. Important regions (e.g., foreground objects, faces) maintain higher quality with less downsampling, while less important regions (e.g., background) undergo more aggressive downsampling. This resolves the contradiction by making quality non-uniform in a controlled, perceptually-optimized manner.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments the video content into different regions of interest and applies region-specific processing. By segmenting the frame and applying different compression levels to different segments based on their importance, the system reduces overall bitrate while preserving quality where it matters most, thus resolving the contradiction between bandwidth reduction and quality maintenance.

Inventive Principle:
Principle #1Segmentation

3Manufacturing precision

If high computational power is used for real-time volumetric video processing, then processing quality is improved, but system complexity and cost increase

Engineering Contradiction:
Improveprocessing qualityVSAvoidsystem complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-training GAN models offline on large datasets of video content. The complex computational work of learning optimal compression and reconstruction strategies is performed in advance during the training phase, rather than in real-time during video processing. This allows the deployed system to achieve high processing quality with reduced real-time computational requirements, thus resolving the contradiction between quality and complexity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by training the GAN on copies of video data and learned representations. The generator network learns to copy and reconstruct video content from compressed representations, capturing essential visual information while discarding redundant data. This copying mechanism enables high-quality reconstruction without requiring proportional computational resources, addressing the quality-complexity tradeoff.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3783895A9Visual quality optimized video compression
Publication Date: 2024.08.07 INTEL CORP
  • EP3783895A9 patent drawingFigure 1~2
  • EP3783895A9 patent drawingFigure 3~4
  • EP3783895A9 patent drawingFigure 5~6

AI summary

Techniques related to providing high perceptual quality video from highly compressed and decompressed reconstructed video are discussed. Such techniques include applying a pretrained decompression upsampling portion of a generative adversarial network to the decompressed reconstructed video to upsample and improve the perceptual quality of the decompressed reconstructed video to generate output video.