Research On Multi-source Image Fusion And Evaluation Method Based On Artificial Neural Perceptual Model

Image fusion techniques can improve the image quality and enrich the information of images, which contribute to research analysis and practical application of image. In this thesis, some researches of multi-source image fusion and fusion quality evaluation methods are achieved. Then some suitable image fusion algorithm models are proposed based on pulse coupled neural network and celluar neural network to implement different types of image fusion. The main work and results are shown as follows:(1) Some methods are proposed for solving the problem of the parameters decision of pulse coupled neural network, simplifying network model and modifying the changing pattern of the neurons thresholds. We use local Laplace energy to decide the couple strength of neurons adaptively. Beyond that, this thesis combines Shearlet transform and pulse coupled neural network for accomplishing the different fusing procedure. The simulation proves that the method this thesis proposes performs well in multi-source image fusion.(2) Through the research of basic theory of celluar neural network, we modify the state equation of the cellular based on the theoretic model. Also, based on the analysis of the stability of the network, this thesis, along with the correlation coefficient and genetic algorithm, has achieved the parameter design of network template and reached a high level performance in different fusion atmosphere, which, has proves the efficiency of the multi-source image fusion method based on celluar neural network.(3) Image fusion quality evaluation methods which are normally used are studied. In this thesis, we choose different objective evaluation methods in different kinds of experiments, included multi-focus images fusion, visible-light and infrared images fusion, and multi-spectrum and full color image fusion. And combined with subjective and objective evaluation, we are able to access to compare different fusion algorithm. The experiments and analysis show that the methods proposed by this thesis can both reach a high level performance in different environments. And the fusion method based on celluar neural network achieves a better result than the regular fusion method.