Image Restoration with Operators Modeled by Artificial Neural Networks

This paper presents a new approach to image restoration based on ANN, considering the learning of the inverse process using a standard image for training under a multiscale approach. Different models of ANN were tested and compared with the traditional techniques. The standard image was artificially degraded to simulate some types of frequent degradation problems. Due to the huge amount of data generated for training the ANN, this paper uses clustering techniques to reduce the training set. The paper proposes a simple restoration method that leads to a sub-optimal solution without the need of prior knowledge estimation of the degradation phenomenon. The ANN based filters were tested with different kinds of degraded images. The mean squared error and the signal-to-noise ratio were used as performance indices to measure the quality of the results of the ANN and of some of the existing methods for comparison. The results show that the ANN based restoration algorithms as proposed in this paper are effective restoration methods. The main advantage of the proposed approach is related to the fact that it does not require an estimation of prior knowledge of the degradation causes for each image.