Mediaanisuodin 16-bittisille kuville
Nguyen, Tam · Aaltodoc (Aalto University) · 2017
This thesis utilizes high-performance computing to filter noisy images by various median filter algorithms. Since smartphones are widely available to consumers, customers take noisy images daily. Thus algorithms to filter noises are needed. Many new cameras support more than 8-bit depth for colors. Using 8-bit filtering algorithms on 16-bit images will typically result in color lost. The 8-bit and 16-bit filtering algorithms are benchmarked in order to observe how the execution changes when the number of bits increases. Median filter algorithms are the focus of this thesis. This work benchmarks the execution time of four median filter methods: constant time median filter (CTMF), naive, improved naive, and median heap. These algorithms are tested by varying the following parameters: image size, color depth, number of threads, filter window size, and different kind of images. If we use only one thread, CTMF outperforms the other methods for large images. However, it is more difficult to exploit multicore processors efficiently in CTMF. With a large number of threads, the other three algorithms will outperform the CTMF algorithm in 8-bit and 16-bit images. However, this is not the case when the filter window size increases and the image size is constant. The execution time of the CTMF algorithm remains steadily lower than others algorithms in 8-bit and 16-bit images.