High Performance Image Registration Using a Distributed Blackboard Architecture

Research student:  Roger Tait
Supervisors:Adrian Hopgood
 Gerald Schaefer
PhD awarded:6 June 2007

Abstract from thesis

Registration is a method used to geometrically align two images taken from different sensors, viewpoints or instances in time. The images are aligned through a combination of translation, rotation, and scaling. A major drawback of registration is the performance burden associated with resampling and similarity calculation. Such bottlenecks limit registration applications where fast execution times are required. In this research, a novel approach to high performance intensity-based registration is presented. Based on a distributed blackboard architecture and implemented as knowledge sources (KSs), a framework called iDARBS (imaging Distributed Algorithmic and Rule-based Blackboard System) provides an underlying worker/manager model. Division of intensity data into segments by a Distributor KS followed by allocation to multiple Worker KSs allows concurrent resampling and similarity calculation to be achieved. The supervision of Worker KS activities, the evaluation of computed similarity, and the optimisation of transform parameters that map between segments are performed by a Manager KS. Conveniently, the modular nature of the approach permits different similarity calculation strategies to be added to the iDARBS framework without change. The successful distribution of intensity correlation and mutual information-based similarity metrics for the alignment of 2D and 3D data captured by a range of sensor types is demonstrated. Experimental results show a speedup factor of three combined with an efficiency of 43% was achieved during image registration using eight Worker KSs. During single-modal volume registration using ten Worker KSs, a speedup factor of seven and an efficiency of 67% was accomplished. Finally, a speedup factor of three combined with an efficiency of 50% was achieved during multi-modal volume registration using six Worker KSs. Crucially, the results reported confirm the success of the approach.

Return to Adrian Hopgood's page.

Valid HTML 4.01! Owner:
  Adrian Hopgood
  March 2019