Digital Image Restoration Algorithms Based On Neural Network

Image filtering and image restoration is a part of image processing. In the past few years, it has been the focus and hot topic of researches. Additionally it occupies a very important position in the digital image processing, which is widely used in satellitical remote sensing images, medical image processing, military, public security and so on. Digital images are often corrupted by noise during image acquisition and transmission due to a number of nonidealities encountered in image sensors and communication channels. Impulse noise is one of the most typical cases. This dissertation studies the image restoration method which mainly deals with the impulse noise.Various image restoration methods have been proposed, such as inverse filter, Wiener filter, maximum entropy, iterative blind deconvolution, etc. However, these traditional methods are inefficient in solving the problem of function approximation. Artificial neural network has its unique advantages in this aspect, because it is a type of large-scale nonlinear dynamical system, characteristic of high-speed parallelism calculation, great robustness, strong capacity of self-adaptive, self-organization and self-learning.Human brain can understand inaccurate and incomplete information, acquired by the perceptional organs. Fuzzy set theory has been proposal for imitated this function of human brain. Because there are a lot of similarities and complementarities between the structure and function of fuzzy system and neural networks, they can be integrated very well to form fuzzy neural network, and broaden their fields of application.Because the traditional impulse noise filtering methods will lead to the degradation of the image details, two neural network-based restoration methods have been proposed in this dissertation. One is based on the RBF neural network, and the other is based on the fuzzy neural network. Because of the advantage of neural network's performance in pattern recognition, we can make a noise detector by a RBF neural network easily. The new restoration method is a filter combined with the noise detector and a median filter, in which the noise detector is able to detect the degraded pixels effectively. So this method can protect the image details. Image restoration method based on the fuzzy neural network is an image data fusion model that using the fuzzy neural network to fuse the edge detection image, the median filtering image and the degraded image. Simulation results show that the proposed two methods can filter the impulse noise effectively, at the same time, can protect the image details very well.