This thesis "Image Transformations on Processor per Pixel SIMD Architectures" is written to obtain the academic degree "Master of Philosophy". It explores the possibilities of geometric image transformations on processor per pixel SIMD architectures. Furthermore, a programming language for the ASPA Vision Chip, which is developed at the University of Manchester, is defined. Cellular single instruction multi data (SIMD) image processing chips like the ASPA hardware architecture combine a processor and a photo sensor for each pixel of a given image grid. The elementary processing cells of such a device are connected within their nearest neighborhood and due to the large amount of processors on these chips, they are capable of calculating complex image processing tasks in real time. This is achieved by using the processing elements (PEs) of the device in parallel and by the utilization of the neighborhood connections. But as this field of research just emerges, many algorithms that are common in classic image processing need to be transferred to this new parallel approach. In this thesis two very general methods of image manipulations are examined: a concept for routing pixel information on the PE grid and an application for a frequency domain image transformation. Both of these concepts deploy a method that uses long range data communication between the processing elements, implemented in an efficient way. These two new methods can be applied easily to image manipulations like zooming, rotating, mirroring and shifting images on the processor array. Results of simulations and implementations of these applications are provided in addition to a detailed description of the above mentioned methods.
|Date of Award||1 Aug 2011|
- The University of Manchester
|Supervisor||Piotr Dudek (Supervisor)|
- Vision Chips, Pixel per processor arrays
- Image transformation, ASPA, Pixel Routing