Hierarchical Image Data Segmentation for Parallel Processing Overhead Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face inefficiencies in parallel processing and increased overhead when dividing image data for execution across multiple computation devices, leading to suboptimal processing efficiency and resource utilization.
Innovation Solution
An image processing device that divides image data into smaller segments based on pre-and-post dependency relationships, allowing for parallel processing across multiple computation units, including CPUs and GPUs, with a control unit managing the execution of these segments to optimize resource allocation and reduce overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image data is divided into division image data and processed in parallel by multiple computation devices, then processing efficiency is improved, but overhead increases due to memory management and task execution
Solution Approach 1:
The patent divides image data into multiple division image data segments that can be processed in parallel by different computation devices. Each division image data is further subdivided into subdivision image data, creating a hierarchical segmentation structure that enables efficient parallel processing while managing memory resources effectively.
Solution Approach 2:
The patent implements a nested structure where subdivision image data (smaller units) are contained within division image data (larger units). This nested doll approach allows the system to manage processing at multiple levels of granularity, enabling fine-grained parallel processing while reducing overall overhead through hierarchical organization.
2Ease of operation
If computation devices use only their own memory for partial processing, then memory management is simplified, but processing efficiency decreases due to increased data transfer requirements
Solution Approach 1:
The patent introduces a storage device as an intermediary between computation devices and the central control unit. This storage device acts as a shared memory resource that computation devices can access, eliminating the need for each device to use only its own memory while maintaining efficient data access and reducing transfer overhead.
3Device complexity
If image data is divided into larger segments, then the number of processing tasks is reduced, but parallel processing efficiency decreases
Solution Approach 1:
The patent applies multi-level segmentation by first dividing image data into division image data, then further subdividing each division into subdivision image data. This creates a hierarchical task structure that increases the number of parallelizable tasks while organizing them in a manageable way through the division-subdivision relationship.
Solution Approach 2:
The patent introduces a hierarchical dimension to the task organization structure. Instead of a single level of segmentation, the system uses two levels (division and subdivision), adding a dimensional layer that enables better parallel processing efficiency while maintaining manageable task complexity.
Data Source
AI summary
An image processing device executes image processing by each object of an object group in which plural objects are connected to each other in a directed acyclic graph form. The image processing device includes: a division unit that divides image data as an image processing target into division image data having a first size; a subdivision unit that subdivides the division image data into subdivision image data having a second size smaller than the first size for each partial processing which is image processing to be performed on the division image data, the division image data corresponding to the partial processing which is determined as executable processing based on a pre-and-post dependency relationship; and a control unit that performs control for causing plural computation devices to execute subdivision partial processing which is image processing to be performed on the subdivision image data, in parallel.