Agent-based, Automatic Parallelization of Image Analysis Tasks The Project-Abstract

Technical References

Location:	Chair Computer Science IX, Technische Universität München, Munich, Germany
Project Basis:	Image Understanding Research Group (FGBV)
Implementation Basis:	Marmot Multi-Agent Library and Halcon Image-Analysis System
Project Integration:	The project is integrated into "Cooperating Agents in Distributed Systems" - a subtask of the project Cooperation and Resource-Management in Distributed Systems
Cooperation:	The project cooperates with the Multi-Agent-Discussion (MAD) working group.
Project Start:	1996-January-01
Project End:	1999-May-01

Motivation of the project

Image analysis applications require large amounts of resources (especially memory) and need an increased performance. Most conventional computing techniques are not able to satisfy these enhanced requirements. On the other hand many new techniques are methodically attached too close to the constraints of a special problem. This restricts portability and the reusing of code.
Thus there is the need of new architecture-independent methods that increase performance of image processing and optimize the resource management.

The Idea

The idea beyond the project is the integration of different methods of parallelism (task-parallelism, data-parallelism, pipelining) within one system for image analysis. By giving it a specification of an image analysis task, the system plans an optimal (parallel) processing of the task. Secondly the plan will be executed by the system. A multi agent system forms the basis of both stages - planning and execution. Thus the agents perform an automatic parallelization of the task and control the scheduling/mapping of subtasks and -data.

The Goals

This approach should show the following advantages:

• optimized resource-usage by parallelism
• advanced performance by parallel processing
• more efficiency by "teamwork": cooperative execution of image analysis tasks and cooperative planning by the multi agent system
• adaption of the whole system via the agent's ability of learning
• easier handling of the system and optimized portability by abstraction of the hardware

Agent Design

This page shows some aspects of the design of our multi-agent system.

Publications

• M. Lückenhaus, W. Eckstein,
  A Thread Concept for Automatic Task Parallelization in Image Analysis, in: Parallel and Distributed Methods for Image Processing II, July 19-20, San Diego, USA, Proc. SPIE 3452, 1998 [Abstract, PostScript Version 0,6 MBytes, compressed (gz) 0,17 MBytes]
  
  
• M. Lückenhaus,
  A Multi-Agent System for Parallelizing Image Analysis Tasks,
  accepted for: International Conference on Intelligent Autonomous Systems 5 (IAS-5), Sapporo, Japan, June 1-4, 1998 [Abstract, PostScript Version 3,9 MBytes, compressed (gz) 1,3 MBytes]
  
  
• Maximilian Lückenhaus, Wolfgang Eckstein:
  A Multi-Agent Based System for Parallel Image Processing
  in: Parallel and Distributed Methods for Image Processing (SPIE's Annual Meeting 97), San Diego, August 28-29, Proceedings of SPIE Vol. 3166, 1997.
  [ Abstract, uncompressed PostScript paper (0.7 MB), compressed PostScript paper (gz: 0.2 MB)]