Neural Reference Motion for High-Resolution Video Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The use of uniformly shaped square coding units in high-resolution images leads to degradation in the quality of reconstructed images, necessitating improved methods for splitting coding units.
Innovation Solution
An image decoding method that generates reference motion information using artificial neural networks, aggregates, scales, or transforms this information, and performs motion compensation on subblocks to enhance the decoding process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If uniformly shaped square coding units are used, then the coding structure is simple and easy to implement, but the quality of reconstructed images degrades in high-resolution images
Solution Approach 1:
The patent divides the coding unit into multiple subblocks (first subblock and second subblock) with different sizes and shapes. The first subblock has size W1×H1 and the second subblock has size W2×H2, where W1+W2=W and H1=H2=H. This segmentation allows each subblock to be optimized independently for motion compensation, resolving the contradiction between structural simplicity and image quality.
Solution Approach 2:
The patent applies different motion compensation methods to different subblocks based on their specific characteristics. The first subblock uses reference motion information from a first reference block, while the second subblock uses reference motion information from a second reference block. This local quality approach allows optimized motion compensation for each region, improving overall reconstructed image quality while maintaining coding efficiency.
2Measurement precision
If reference motion information is generated using artificial neural networks, then the motion information accuracy is improved, but the processing complexity and computational resources increase
Solution Approach 1:
The patent performs preliminary actions by generating reference motion information for reference blocks before processing the current block. The neural network generates reference motion information for multiple reference blocks in advance, which are then stored and used during the motion compensation process. This preliminary action reduces the computational burden during actual block processing while maintaining high motion information accuracy.
Solution Approach 2:
The patent introduces an intermediary approach by using a neural network as a mediator between the input image and the motion information generation process. The neural network processes pixel data of reconstructed pictures to generate motion information, acting as an intermediary that transforms raw pixel data into accurate motion vectors, thereby improving measurement precision while managing processing complexity through specialized computation.
3Manufacturing precision
If motion information is obtained from multiple reference blocks, then the motion compensation accuracy is improved, but the number of processing steps and data handling complexity increases
Solution Approach 1:
The patent segments the current block into multiple subblocks, with each subblock associated with specific reference blocks. The first subblock uses motion information from the first reference block, while the second subblock uses motion information from the second reference block. This segmentation organizes the complex task of obtaining motion information from multiple reference blocks into manageable, independent operations, reducing data handling complexity while maintaining compensation accuracy.
Solution Approach 2:
The patent applies local quality by selecting specific reference blocks for each subblock based on their spatial relationships and motion characteristics. Each subblock receives motion information tailored to its specific region, allowing optimized motion compensation for each area while managing overall data handling complexity through localized processing.
Data Source
AI summary
An image decoding method includes generating reference motion information for at least one block included in a current picture, based on at least one of: pixel data of at least one reconstructed picture, or motion information of the at least one reconstructed picture; obtaining motion information for a current block from the at least one block included in the current picture, based on the reference motion information and at least one syntax information obtained from a bitstream; and reconstructing the current block, based on the motion information for the current block. The method may include generating the reference motion information using a neural network. An image encoding method includes similar steps of generating reference motion information, determining motion information for a current block, and encoding the motion information into a bitstream, based on the reference motion information.