Study On Digital Image Pre-processing And Fusion Methods

With the development of sensor technique, the methods that people obtain imageshas been becoming more and more, making the category of image processing also moreand more. The classic methods have not done it well. It needs to develop new methodsto fusion images, while intelligence calculation cognition-based that is developing fastrecently has found out good solution. The objective of this thesis is to propose someeffective methods to fusion images combining neural network, genetic algorithmsstatistic theory and so on, after discussing some problems about pre-processing.This research mainly has two parts: the first part is the pre-processing of imagefusion, including image filter, image segmentation, and image matching and the secondpart is about the research of image fusion algorithm. This research is mainly around thewhole processing of image fusion, and obtains research results as follow.(1) The effect by common image restoration method is not good. Then a method ofimage restoration is designed according to Hidden Markov tree and wavelet transform.The image is decomposed using double-tree complex wavelet after Wiener filtering.And noise is removed based on Hidden Markov tree in complex wavelet domain. Theresults show that the new restoration algorithm is more effective than existing methods.(2) To overcome the shortcomings of large computation and complicated modelrepresentation in image segmentation based on deformable model, a novel segmentationmethod is proposed based on across entropy and curve evolution. Across entropy isutilized to construct energy function from the difference among the classifications, thesegmentation is transformed to minimize the energy function. Curve evolution equationis achieved by the energy function, and level set is employed to represent and solve theequation for effective image segment.(3) Aim at the precision and speed of image registration, an algorithm is improvedbased on wavelet transform and mutual information. The layered match scheme isadopted to improve the matching speed. Probability density is estimated with Parzenwindow. The optimum metric of mutual information is solved using gradient method.So the precision and speed of image registration is improved. This method can beapplied to affine between current image and reference image.(4) A new lift scheme is designed in adaptive wavelet transform to overcome thecomplicated problem of wavelet reconstruction. The new algorithm can carry out thecomplete reconstruction without other additional information very naturally. By averaging the approximate image signal and choosing the bigger detail image signal tobe fusioned, this algorithm is applied to image fusion. The imitative result shows thatthe method can obtain detail component and be robust in multi-sensors fusion.(5) A new algorithm is proposed based on fuzzy kernel clustering and neuralnetwork to overcome the problem which happens in algorithm of image fusion byneural network.. The algorithm adopted wavelet transform to filter the mixture noiseimages and design fuzzy kernel clustering to improve the result and speed. The resultsshow it effective comparing to other image fusion algorithm based on neural network.(6) A new algorithm is proposed based on genetic evolution algorithms and fuzzyneural network to solve problem of the structure and the parameters which exists inimage fusion algorithm based on neural network. The method improves the effectivefrom signal-to-noise through making fuzzy neural network as clustering in single image,the max signal-to-noise fusioned image as fitness function, adjusting the innerparameters of fuzzy neural network. The results show it effective comparing to othermethod based on SOM network and fuzzy neural network.(7) The robust of traditional image fusion algorithm is bad. So a novel imagefusion algorithm is designed based on EM algorithm. This method infers the iterativeprocess of real scene to fusion images according to EM algorithm based onmulti-images model. Varies of methods, such as Laplace Pyramids, gradient pyramids,wavelet, wavelet frame and direction filters Pyramids, are discussed and compared toEM algorithm. The digital imitation results show that the robust of the method is quitegood even when fusion the seriously polluted image.