An Evaluation of MLP Neural Network Efficiency for Image Filtering

Digital images have been widely used in a diversity of areas. The processes of digitalization and transmission of images may reduce its visual quality. Many recent publications include different tools to improve the quality of these images such as filters that reduce the noise within the images. Neural networks, such as multilayer perceptrons (MLP), have been widely studied as efficient tools for arbitrary function approximation. This paper describes the usage of MLP working as filters for images. More specifically, the work deals with the reduction of impulsive noise in images and describe conditions to obtain efficient results in comparison with the traditional filters. A simply method to evaluate the efficiency of filters is described. The factor considers not only the ability of the filter in restoring noisy pixels but also the effects that may occur in original (noiseless) pixels, such as blurring, that degrades the images. Experiments were conducted for traditional filters, such as mean and median, and MLP networks with increasing presence of noise within the training data. Results shown a better overall efficiency for MLP networks trained with images degraded with a small percent of impulsive noise.