Home»Computers-and-Technology»Solving Large-scale Image Processing Computations: What is the Most Suitable Architecture?

Solving Large-scale Image Processing Computations: What is the Most Suitable Architecture?

Add to Favorites

Send to a Friend

Publish this on your site

Send to a friend by YM

Author: Simon Loader

This article gives advice on complex computer architectures to be used for solving large-scale image processing computations in a relatively fast response time. Considering the various complex architectures, the architecture belived to be the most suitable type is outlined with its essential characteristics.

Large scale image processing computations usually require large matrix operations to be made across many images or for the same process to be done over a large collection of images/data. For this reason it is considered this type of problem is suitable for SIMD and MIMD type machines as these allow the manipulation of multiple data items and give a relatively fast response.

The best suited SIMD type would be distributed memory SIMD also known as array processors. This would allow each processor to contain data for part of an image ready to be processed as required. These types of machines provide fast calculations of matrixes and data manipulation over a large data set just as required for image processing. They are often design to be specialist problem solvers depending on the interconnection network of all the processing elements. For this type of problem solving planner mesh’s have been the most popular and are probably best suited for this problem.

On the other hand MIMD may seem to be too complex as multi instruction should not be required but not if used in a SPMD approach (Single Program Multiple Data). Again this would be using distributed memory in a planer mesh type interconnect but would allow the image processing slightly more intelligence. Instead of maybe having to execute a number of instructions to manipulate all the data a single program can be written to execute on each processing element. This should be able to give a slightly higher average turn around as with SPMD some data sets will not need the operation perform and they can be skipped unlike with a SIDMD processors.

I would recommend the use of a MIMD type architecture designed to run a SPMD. This would allow greater flexibility in the future. It may also lead to the use collection of workstation as local processor speeds increase and therefore reduce costs. MIMD architecture can mirror all others of Flynn’s classification just by how they are used and may prove useful for other computations in the future. The relative costs are liable to be the same as MIMD tend to use generic processors compare to the more specialised SIMD architectures.