Pixel Level Image Fusion Technology

Image fusion is an important branch of data fusion, whose goal is to achieve a comprehensive description about the same scene by combing some source images obtained from different sensors. So the fused image contains more useful information that is more convenient than source images in the consequent processing, analysis and interpretation. Now, image fusion is widely applied in many fields such as medical image analysis, remote sensing, military application, information encryption, etc.Image fusion can be divided into three levels, i.e. pixel-level fusion, feature-level fusion and decision-making level fusion. Being the foundation of the other two levels fusion, pixel-level image fusion can provide with abundant, accurate and reliable information that doesn't exist in the other levels. So it is not only the most complicated one, but also the most difficult one to implement. This dissertation focuses on pixel-level image fusion and the major contributions are as follows:(1) The principles and steps of pixel-level fusion are analyzed first. Then some standard methods on pixel-level fusion are summarized in details after theirs advantage and disadvantage especially their potential applications are discussed.(2) Performance evaluation on image fusion is discussed and some widely-used quality parameters are given that include subjective and objective methods. Then a standard of selecting evaluation is suggested. At last, a trend of image fusion evaluation is proposed, i.e. developing a comprehensive evaluation system of image fusion which is integrated with subjective items and objective ones.(3)A method on image fusion based on the optimal comprehensive performance via wavelet transform is presented. First, according to the fusion purpose and experts' experience, some objective evaluation standards of fusion are selected and used to construct a comprehensive evaluation model. Second, each source image is decomposed by wavelet transform with the fusion parameters generated from the initialization of Genetic Algorithm (GA), and then several fusion images and their comprehensive performance values are computed. Next, guided by the fitness function representing the optimal performance, GA approaches the best performance parameters generation by generation via operations such as selection, crossover and mutation. After the optimal parameter is located, the best fusion image can be obtained by the corresponding wavelet reconstruction. Finally, experiment results show that our method is superior to many existing fusion schemes.(4) In terms of the characteristic of IR images, we present a method on object extraction via grey theory, whose idea is to detect and extract some useful object information from infrared image by grey relational analysis theory. Based on the method, we propose a new image fusion method to fuse infrared and visible images.In the method, we first analyze the features of infrared and visible images respectively; and then acquire the final object directly from the infrared source image. Some experimental results indicate that our fused effect has been obviously improved, and is superior to some traditional algorithms like CP(contrast pyramid) method, for that our fusion images not only possess the same clear object as the one in infrared image, but also keep the same detail and background as the ones in visible image.