Skip to Main content Skip to Navigation

Architecture dédiée au traitement d'image base sur les équations aux dérivées partielles

Eva Dejnozkova 
Abstract : The image processing methods based on Partial Differetial Equations (PDEs) draw a groving attention of the scientific community. The number of applications has considerably increased with the introduction of the Level Set methods. The EDPs apply in many domains such the image improvement (non linear diffusion), the segmentation by active contours used both for static pictures (graphe of Voronoï, watershed, shortest paths, object detection) and image sequences (object tracking) or more recent methods such as shape-from-shading. The industrial applications of these methods remain very limited due to a considerable computation complexity (high iteration number) on one hand, and the difficulties of embedded system implementation (energy consumption, memory requirements) on the other hand. To our knowledge, few real time experiments on supercomputers and graphics hardware were published, and embedded applications are almost inexistent. Our goal is to propose a dedicated architecture facilitating the implementation of a real time embedded system. Regarding this objective, we propose a new algorithm (called Massive Marching) solving the Eikonal equation for computation of the distance function. It proceeds in parallel and eliminates the usage of the ordered data structures. It allows to obtain the solution either on the entire image or only on its part: the Narrow Band around the propagation front. The complexity of Massive Marching is linear. We consider that its introduction is important also for the community of Mathematical Morphology, because Massive Marching represents the algorithm allowing to obtain the watershed in parallel. Next, we propose two architecture types (i) SIMD and (ii) MIMD, based on several embedded processor cores, implementing Massive Marching in parallel or semi-parallèle. The same architecture types can be used to implement filtering algorithms as well as methods of interface evolution. The same architecture can therefore be used to implement all the steps of a complete application consisting of different types of algorithms, e.g. filtering followed by segmentation.
Document type :
Complete list of metadata
Contributor : Ecole Mines ParisTech Connect in order to contact the contributor
Submitted on : Tuesday, April 5, 2005 - 8:00:00 AM
Last modification on : Saturday, March 31, 2018 - 11:04:01 AM
Long-term archiving on: : Thursday, September 30, 2010 - 6:26:36 PM


  • HAL Id : pastel-00001180, version 1



Eva Dejnozkova. Architecture dédiée au traitement d'image base sur les équations aux dérivées partielles. domain_other. École Nationale Supérieure des Mines de Paris, 2004. English. ⟨pastel-00001180⟩



Record views


Files downloads