Abstract:
Quadratic filters are usually more successful than linear filters in dealing with nonlinear noise characteristics. However, determining the proper weights for the success of quadratic filters is not straightforward as in linear case. For this purpose, a search algorithm used to train weights of quadratic filters from sample images by formulating the problem into a single objective optimization function. In the presented study, comparative inspections for training quadratic image filters using genetic algorithm (GA) and particle swarm optimization (PSO) were presented. Because computation of fitness function involves consecutive image filtering operation using candidate solutions, this process usually results in long training durations due to the computationally expensive nature of image processing applications. In order to reduce the computation times, variable and variable random fitness methods were implemented, where the image size varied in the computation of fitness function. Experimental results show that proposed algorithm provides about 2.5 to 3.0 fold acceleration in computation durations using both GA and PSO.
