Gies, Valentin (2005) Asynchronism in artificial retinas Thesis, ENSTA.

Restricted to Repository staff only



Image processing is usually divided into low- mid- and high-level operations. Mid-level regional operators are a key feature in image processing. In a preliminary study, we present their importance considering image processing with an energetic point of view. We also determine a general way of implementing these operators in order to use them in an efficient way. After having the necessity of asynchronism for implementing regional operators, we review existing structures for massively parallel regional computation. Because the hardware cost of asynchronism is too important for a dense implementation in artificial retinas, we propose to reduce to the minimum possible the hardware ressources needed for regional asynchronous computations. This leads to a new regional computation structure, the associative micropipelines, implementing part of the associative mesh model, especially the regional sum. This structure is based on an asynchronous element, the convergent micropipeline having a reduced hardware cost. The associative micropipelines use an alternative mode of asynchronous communication we call the token transmission. In order to reduce again the regional operators hardware cost defined before, we propose a way to optimize the asynchronous interconnexion network. In the algorithmic part, we show that the possibility of using efficient regional operators allow to extend the image processing framework. We propose to illustrate this extension applying it to image segmentation. After a review of the existing segmentation methods, we propose a statistical solution to watershed over-segmentation issue. Finally, we present a novel segmentation method, societal segmentation, based on an analogy between segmentation of a land into towns and villages and image segmentation, using regional measures in an intensive way.

Item Type:Thesis (Thesis)
Uncontrolled Keywords:Segmentation
Subjects:Information and Communication Sciences and Technologies
ID Code:1662
Deposited By:Julien Karachehayas
Deposited On:04 avr. 2006 02:20
Dernière modification:05 juin 2013 09:13

Repository Staff Only: item control page