Image spatial filtering using real development platform for advanced driver assistance systems

U ovom radu implementirani su i uspoređeni različiti tipovi prostornih filtara. Najprije su predstavljeni osnovni tipovi šuma koji stvaraju potrebu za filtriranjem, a to su: Gaussov, šum „soli i papra“ te kvantni. Nakon toga su predloženi načini suzbijanja istih različitim tipovima filtara te je dan teorijski osvrt na četiri prostorna filtra: medijan, moving average, Gaussov i visokopropusni. Razvijeni algoritmi implementirani su na realnu razvojnu ADAS platformu s prilagođenim C kodom za Alpha ploču i kreiranim slučajevima upotrebe za svaki filtar za 3 različita rasporeda procesora: DSP, A15 i kombinacija dvaju DSP procesora. Dodatne kombinacije filtara i procesora su filtri s različitim veličinama kernela i različitim rezolucijama. Svaka od navedenih kombinacija se testirala na testnom skupu od deset slika. Rezultati su uspoređeni te je utvrđeno da se najbrže izvodi visokopropusni filtar, a najsporije medijan. Ovisno o primjeni, može se koristiti veća rezolucija, ali ona, kao i povećanje veličine kernela, uzrokuje smanjenje brzine izvođenja.In this paper were implemented and compared different types of spatial filters. Firstly, the basic types of noise, which create the need for spatial filters, are described. Common types of noise are: Gaussian, „salt and pepper“ and quantization noise. After that are named and described different types of spatial filters used for reducing noises in digital image. These filters are: median, moving average, Gaussian and highpass filter. The developed algorithms are implemented on a real ADAS development platform with custom C code for Alpha board and created usecases for each filter for three different CPU-s: DSP, A15 and a combination of two DSP CPU-s in parallel. Aditional combinations are three different resolution and three different kernel sizes in each filter. Each of this combinations were tested on a test set of ten images. The results were compared and it was found that high pass filter is the fastest, and median the slowest. Depending on the application, a higher resolution can be used, but it, as well as increasing the kernel size, causes a decrease in execution speed