-In theory, modern graphical processing units (GPUs) make parallel programming accessible to all, and have triggered widespread interest among researchers or developers of all disciplines, with the hope of dramatically increasing processing speeds. Nevertheless, obtaining such processing speeds as expected cannot be done without considerable designing efforts : as an answer, we propose in this thesis, two GPU-based methods leading to fast implementations of several algorithms targeted to processing noisy images. One of them consists in porting the segmentation algorithm named \textit{snake}, with the effect of extending its processing capacity and performance. A second involves an innovative GPU-specific algorithm, based on searching for level lines within gray-level or color images to reduce gaussian noise, whose signal-to-noise ratio is particularly interesting.
-Through extremely fine-tuned management of the different memory types available on GPUs, we have also conferred unprecedent flow rates to the median filter, making it able to process over 5 million pixels per second. Eventually, we extended the above methods to the more generic convolution filter, and showed they out-perform the fastest implementations known to date, with over 7 million pixels per second. In addition, we provide an on-line application that enables any developer to automatically generate operational source code of our filters.
+In theory, modern graphical processing units (GPUs) make parallel programming accessible to all, and have triggered widespread interest among researchers or developers of all disciplines, with the hope of dramatically increasing processing speeds. Nevertheless, obtaining such performances cannot be done without considerable designing efforts : as an answer, we propose two GPU-based methods leading to fast implementations of several algorithms targeted to processing noisy images. One of them consists in porting the segmentation algorithm named \textit{snake}, with the effect of extending its processing capacity and performance. A second involves a innovative GPU-specific algorithm, based on searching for level lines within gray-level or color images to reduce gaussian noise, whose quality-to-speed ratio is particularly interesting.
+Through extremely fine-tuned management of the different memory types available on GPUs, we have also conferred unprecedent flow rates to the median filter, making it able to process over 5 billion pixels per second. Eventually, we extended the above methods to the more generic convolution filter, and showed they out-perform the fastest implementations known to date, with over 7 billion pixels per second. In addition, we provide an on-line application that enables any developer to automatically generate operational source code of our filters.
