Researches On Premature Convergence In The Genetic Algorithm And Its Application In Image Restoration

In the process of image forming, recording, scanning, transferring, showing, etc, the presence of image degradation is unavoidable. However, in many application areas, the clear images of high quality are needed. The image restoration, aiming at restoring the original image from the degraded one, is fundamental to image processing, pattern recognition, computer vision and is widely used in astronomy, remote sensing, medical imaging, etc.In image restoration, because of the complexity and close correlation of the image information, at different levels in image processing maybe some problems arise, such as non-integrity, non-precision, non-structure, etc. In conjunction with computational intelligence methods, image restoration can get a better effect in comparison with the traditional image restoration methods.The genetic algorithm (GA) as an intelligence optimization method makes use of natural selecting and genetic rules for stochastic searching. The primary character of GA is that it can operate synchronously at multiple points in the searching space for the global optimization. The way of searching is by the stochastic rules and is independent of the gradient information. So GA is capable of solving the complexity and nonlinear problems, which may cause difficulty for the traditional methods. The predominance of image restoration based on GA can find a solution from non-integrity, non-precision, and partial facticity information with lower cost. Consequently, this method has great potentialities for further development of image restoration technology.This thesis focuses on premature convergence in GA and its application in image restoration. The three aspects of the work done are shown as follows:(1) An improved GA is proposed to overcome premature convergence of the GA. Firstly, the proposed algorithm uses a population diversity operator, which is designed from two aspects-gene and individual, to initialize population with better distribution and to judge whether premature convergence occurs. Once premature convergence appears or tends to appear, the catastrophe operation is implemented to renew population evolution of the algorithm. Secondly, the population passed through selection and crossover operator of the algorithm is made up of two aspects. One is preserved population from parent population, the other is crossover population which is produced by cross of best individual and introduced random population. Preserved population and crossover population constitute anew the next population. Experimental simulation and comparison with standard genetic algorithm (SGA) and elitist genetic algorithm (EGA) demonstrate that the proposed algorithm in function optimization can maintain population diversity and find overall optimum solution.(2) Aiming at premature convergence of the GA in image restoration, an improved genetic algorithm in binary image restoration is proposed. The algorithm introduces random population with proportion r every k generation, in which r is random population proportion in population size n. If introducing random population in some generation, the crossover operator of this generation is the crossover between current best individual and every individual which is in the population after joined random population. At the same time, the improved algorithm designed a local mutation operator only to best individual. Toward anyone mutation point, Canny operator should be used to inspect whether there is edge information in the local region of the mutation point and its eight pixel-neighbors. For the locally connected region without edge information, the pixel value is taken as 0 or 1 or the original value of the locally connected region is unchanged. Every certain generation introducing random population once strengthens the algorithm's ability of searching new solution space, and the local mutation operator based on edge information quickens convergence to best individual. Experiments demonstrate that the proposed algorithm is better than SGA and can be very insensitive to blur level of the image.(3) In image restoration, noise reduction is also one of important research aspects. Through research of L-filter, a novel algorithm for mixed noise filtering in image processing is presented. The algorithm based on central limit theorem estimates the mixed noise model through inter-selecting region of interest in the image, and adds this mixed noise model to a small test image for rebuilding degraded process. Aiming at this test image, the GA is used to optimize the weight coefficients of L-filter. Then the optimized weight coefficients are used in combination with image edge information to execute L-filter to the image. In L-filter, the filter point whose filter window is with edge information should not be performed. Because Gaussian noise and impulse noise are widely present and are of typical representativeness, the main task of our work is to study how to filter the mixed noise of Gaussian noise and impulse noise. Experiments show that the proposed algorithm is better than Laplacian filter and median filter.