Evolved Filter Design Of Image Denoising

As the computer technology and communication technology developped with high speed, digital image has become an important source to obtation information. With greater amount of information and faster transmission speed, the digital image technology has been a hot and difficult issue of great concern. In the process of image transmission and acquisition, images are inevitably polluted with a variety of noises from the external environment and the physical devices. In order to improve the image quality, image denoising is the pre-processing steps for the follow-up research.Based on the above idea, this dissertation is dedicated to the filter design with evolve algorithm. The emphasisz are noise analysis and filter designs both of gray images and color images.The main characters of this dissertation are as follows:(1)Traditional weighted median filter model is formed based on single dimension signal processing. Each pixel under the template has only one weight. A novel weighted median filter model with more adjustable weithts parameters is proposed here, which considers different correlations among the pixels of the template. Differential evolution algorithm is adopted for the weights optimization. Experiments show that this improved weighted median filter hans better properties both on noise filtering and details protecting.(2) Based on the model of above improved weighted midian filter, correlations between pixel and other pixels under the filtering template are considered. In order to keep stability of filtering on rotated and reversed images, symmetry of the correlations is utilized to form a new filter model. Experiments show that this rotation invariant model exhibits faster convergence characteristics and properties of efficient noise attenuation.(3) Another direction of evolved filter design is also considered. This evolved filter employs genetic programming to evolve the structure of the filter and remove complicated noise. An incremental evolution strategy is proposed for this large training samples evolution mechanism. The amount of sample data is gradually increased by stages. Experiments show that the incremental evolved filter reduces the algorithm complexity and gained better properties of nose attenuation. (4) Also the stability of the rotated and reversed images is considered in the structure design of evolved filter. Based on the analysis of the structure symmetry, a new model with rotation invariant is proposed. Experiments on rotated and reversed images show that it has robust stability.(5) A simple demo system of evolved filter is realized. It can load both gray and color images, input the noise density, set the parameters of evolution algorithm. The evolved filter is compared with other filters. The filtering performance and the filtered images can been showed in this system.