Image Processing Memory Bandwidth Reduction via Unified Reference Block Storage

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video processing technologies require high memory bandwidth to store and retrieve reference image blocks for motion compensation, leading to increased system costs and memory bandwidth demands, especially when dealing with high-definition video encoding and decoding.

Innovation Solution

The method determines a location range in the reference image that covers all or some reference image blocks based on the correlation between motion information, allowing for the storage and retrieval of pixels within this range, reducing the need for separate storage of each reference image block and thus lowering memory bandwidth requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If multiple reference image blocks are stored separately for motion compensation, then motion information accuracy is improved, but memory bandwidth requirement increases

Engineering Contradiction:
Improvemotion information accuracyVSAvoidmemory bandwidth requirement
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges multiple reference image blocks that are spatially adjacent into a single unified storage region. Instead of storing each reference image block separately in memory, the method combines them into one continuous memory area, reducing the total memory bandwidth required for reading and writing operations while maintaining access to all necessary reference blocks for motion compensation.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified storage region serves multiple functions simultaneously: it stores multiple different reference image blocks that may be needed for different motion compensation operations, and it provides a single memory access interface that can serve multiple prediction blocks. This multi-functional design reduces memory bandwidth requirements while maintaining the ability to support various motion estimation and compensation tasks.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If multiple reference image blocks are stored separately, then motion compensation quality is maintained, but system cost increases

Engineering Contradiction:
Improvemotion compensation qualityVSAvoidsystem cost
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges multiple reference image blocks into a unified storage region, reducing the complexity of memory management hardware and software. By consolidating storage requirements, the system needs fewer separate memory buffers and associated control logic, thereby reducing overall system cost while maintaining the ability to perform high-quality motion compensation.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent changes the memory access pattern from multiple separate reads/writes to a single unified read/write operation. This parameter change in memory management reduces the computational overhead and hardware resources required, lowering system cost while preserving motion compensation quality through efficient access to all necessary reference blocks.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If separate storage is used for each reference image block, then access flexibility is improved, but memory bandwidth consumption increases

Engineering Contradiction:
Improveaccess flexibilityVSAvoidmemory bandwidth consumption
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent merges multiple reference image blocks into a unified storage region that maintains internal organization for efficient access. The unified region allows the system to access any specific reference block within it while reducing total memory bandwidth consumption, as the merged structure enables more efficient memory transfer operations compared to accessing multiple separate blocks.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

While merging blocks into a unified storage region, the patent maintains logical segmentation within that region, allowing the system to access specific reference blocks as needed. This segmentation within unity provides access flexibility similar to separate storage, while the unified external structure reduces memory bandwidth requirements.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3068137B1Method and device for image processing
Publication Date: 2021.04.07 HUAWEI TECH CO LTD
  • EP3068137B1 patent drawingFigure 1
  • EP3068137B1 patent drawingFigure 2a~2b
  • EP3068137B1 patent drawingFigure 3a

AI summary

An image processing method and apparatus are provided, which can reduce a requirement on memory bandwidth. The method includes: acquiring N pieces of motion information from N adjacent image blocks adjacent to a current image block, where the N adjacent image blocks are in a one-to-one correspondence with the N pieces of motion information, the N pieces of motion information are used to indicate N reference image blocks in a reference image of the current image block, and the N pieces of motion information are in a one-to-one correspondence with the N reference image blocks; determining candidate motion information from the N pieces of motion information according to a preset rule; determining, in the reference image, a location range of a to-be-stored pixel, and storing all pixels in the location range, where the location range covers all pixels of a candidate reference image block, the candidate reference image block is at least one image block of the N reference image blocks, and the candidate motion information is motion information corresponding to the candidate reference image block; and reading the pixels in the location range, and performing encoding processing on the current image block according to the pixels in the location range.