Evolutionary design of robust noise-specific image filters

Evolutionary design has shown as a powerful technique in solving various engineering problems. One of the areas in which this approach succeeds is digital image processing. Image filtering represents a wide topic in 2D signal processing. In this case different types of noise are considered in the filtering process to restore the image quality that has been decreased by changing values of some pixels in the image (e.g. due to the transmission through unreliable lines or in the process of acquiring the image). Impulse noise represents a basic type of non-linear noise typically affecting a single pixel in different regions of the image. In order to eliminate this type noise median filters have usually been applied. However, for higher noise intensity or wide range of the noise values this approach leads to corrupting non-noise pixels as well which results in images that are smudged or lose some details after the filtering process. Therefore, advanced filtering techniques have been developed including a concept of noise detection or iterative filtering algorithms. In case of the high noise intensity, a single filtering step is insufficient to eliminate the noise and obtain a reasonable quality of the filtered image. Therefore, iterative filters have been introduced. In this paper we apply an evolutionary algorithm combined with Cartesian Genetic Programing representation to design image filters for the impulse noise that are able to compete with some of the best conventionally used iterative filters. We consider the concept of noise detection to be designed together with the filter itself by means of the evolutionary algorithm. Finally, it will be shown that if the evolved filter is applied iteratively on the filtered image, a high-quality results can be obtained utilizing lower computational effort of the filtering process in comparison with the conventional iterative filters.