Investigation For Key Issues Of Image Fusion

Image fusion is the process of combining relevant information in two or more images or sequence information of a scene delivered by different sensors simultaneously or asynchronously into a single highly informative image. The major objective of image fusion is reducing the uncertainty of images. It is realized by increasing the reliability by dealing with the redundant information among multi-images and improving the definition by synthesizing the complementary information among multi-images. Image fusion has become the focus of many research fields recent years.The emphasis in this dissertation is placed on pixel-level image fusion and some new concepts, methods and approaches have been put forward based on the intensive study of several key issues involving image filtering, optimization calculation, image registration and image fusion. In addition, computational intelligence technologies, including rough sets, particle swarm optimization and wavelet analysis, are used in image filtering, registration and fusion. This instructive attempt will redound to enabling image fusion technology to develop in the direction of automatization, universality and intelligence.Image filtering is the pretreatment step of image fusion and has an important influence on the precision of image registration and the effect of image fusion. Two new filtering methods for pulse noises are proposed in this dissertation. The first method is an adaptive filter based on double-window and extremum compression. It integrates some novel strategies, such as separating detecting window and filtering window, extremum compression and adaptive filtering between median filtering and extremum filtering, to improve the detecting veracity and the filtering validity. The second one is based on rough sets theory and difference image. It adopts rough sets theory in dividing the image into steady noise sub-image, pulse noise sub-image and normal pixel sub-image according to the number of directions obtained from difference image, and then introduces different methods to filter the noises in different sub-images. The result shows that these two new methods have better filtering performance than median filter and its improved methods.Optimization calculation plays an important role in image registration and fusion. An improved Particle Swarm Optimization (PSO) algorithm is proposed in this dissertation by eliminating the paticle velocity variable and adding an extremum trembling operator. The particle velocity variable, analysised and proved in theoretics, is unnecessary. So the evolutionary equation of PSO is rewritten without the particle velocity variable. The new PSO is more simple but efficient, and its differential equation is decreased from second order to first order. In the meantime, an extremum trembling operator is designed for PSO to escape from local extremum after the reason is drawn from the characteristic of evolutionary equation. The results indicates that the PSO without particle velocity variable improves greatly the precise and velocity of convergence, the PSO with the extremum trembling operator escapes efficiently the local extremum, and the PSO combined the two strategies obtains the better optimization results with smaller population size and evolution generations. The new algorithms improved the practicality of the particle swarm optimization.Image registration is an important step of image fusion and its aim is eliminating or decreasing the difference of image in time, space, phase and resolutions. A new method based on wavelet decomposition and particle swarm optimization is put forward for rigid image registration. Image registration in this method is divided into coarse registration phase and refined registration phase. During the coarse registration phase, the improved particle swarm optimization is adopted to perform the registration of the low-frequency parts decomposed by wavelet analysis from two images using mutual information as the registration measurement. During the refined registration phase, variable alternation is adopted to perform the quick search process for the optimum in the small adjacent domain based on the result of coarse registration. The result indicates that this new method has such advantages as antinoise performance, high precision, and high success rate of registration and it is a universal and automatic registration method for rigid images. These virtues endow it with promising applications.The kernel of image fusion is designing fusion rules for calculating the fusion coefficient of pixels corresponding to the positions in the images. The pixel value and local eigenvalue are often used to calculate the fusion coefficient. A new method based on subtractive image segmentation for weighted image fusion is put forward from another viewpoint. It adopts the improved particle swarm optimization to calculating the segmentation threshold of subtractive image and the corresponding weighted fusion coefficient in space domain or in wavelet decomposition domain. Moreover, it doesn't need to construct the eigenspace and quantitative eigenvalue for the weighted fusion coefficients. The result shows that this method has complete automatization and good performance in image fusion.In summary, this dissertation makes exploring researches on image filtering, optimization calculation, image registration and image fusion using computational intelligence technology. New methods are put forward based on the theoretical basis and the anticipative experimental results, and are beneficial to the theoretical research and engineering applications in image fusion technology. It is worth to mention that the improved particle swarm optimization shows good applicability and can be used in other fields concerning optimization.