Video Compression Using Pixel Kernels for Accurate Motion Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression algorithms using optical flow networks for motion estimation suffer from inaccurate pixel-level offset prediction, leading to additional residual compensation information due to complex motion deformation, which complicates video storage and transmission.
Innovation Solution
Generate a pixel kernel for each pixel in a key frame based on a key frame and a forward search frame, perform smoothness constraint processing to obtain a target pixel kernel, compress the key frame and target pixel kernel, and compress the residual graph between the forward search frame and a predicted frame to obtain a compressed video packet.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If optical flow network is used for pixel-level offset prediction, then motion estimation can be performed, but prediction accuracy deteriorates in complex motion deformation cases
Solution Approach 1:
The patent segments the motion estimation process into multiple levels: first obtaining coarse motion information through optical flow network, then refining it through pixel kernel processing and smoothness constraints. This multi-level segmentation allows the system to handle complex motion deformations by progressively improving prediction accuracy at each stage.
Solution Approach 2:
The patent introduces pixel kernels as an intermediary element between the optical flow network output and the final motion compensation. These pixel kernels act as mediators that capture local motion characteristics and enable more accurate prediction for complex deformations without requiring complete redesign of the optical flow network.
2Measurement precision
If additional residual compensation information is added to improve prediction accuracy, then motion estimation quality improves, but data amount and complexity increase
Solution Approach 1:
The patent changes the parameter representation from traditional residual compensation to pixel kernel parameters with smoothness constraints. By transforming the motion representation into a different parameter space (pixel kernels), the system achieves better prediction accuracy without proportionally increasing data amount, as the kernel parameters can be efficiently compressed.
Solution Approach 2:
The patent performs preliminary smoothness constraint processing on pixel kernels before final compression and transmission. This preliminary action reduces the complexity and data amount of the motion information beforehand, making the subsequent compression more efficient and reducing the overall data requirement for accurate motion estimation.
3Ease of manufacture
If traditional compression methods are used, then implementation is simple, but compression efficiency is insufficient for large video files
Solution Approach 1:
The patent replaces traditional mechanical compression approaches with AI-based processing. By substituting conventional compression mechanics with neural network-based pixel kernel generation and smoothness constraint processing, the system achieves superior compression efficiency for large video files while maintaining implementability through standardized AI processing pipelines.
Data Source
AI summary
This application relates to a video compression and video decompression that includes: generating, based on a key frame and a forward search frame of a video, a pixel kernel of each pixel in the key frame; performing smoothness constraint processing on the pixel kernel, to obtain a target pixel kernel; compressing the key frame and the target pixel kernel, to obtain a compressed key frame and a compressed pixel kernel; compressing a residual graph between the forward search frame and a predicted frame, to obtain a compressed graph, the predicted frame being a video frame generated based on the target pixel kernel and the key frame; and obtaining a compressed video packet according to the compressed graph, the compressed key frame, and the compressed pixel kernel.