Applications Of Pulse-Coupled Neural Networks In Image Registration And Interpolation

With the rapid development of science and technology in the last half century, artificial intelligence(AI) technology has become one of the mainstream techniques and adopt itself deeply into human's living and production. Artificial neural network(ANN), the artificial models of biological nerve system, which can simulate the intelligence of human's brain to some extent, has shown its great vitality and wide application prospect in the AI technology. Pulse-Coupled Neural Network (PCNN) is the third generation of ANN models, which has many excellent characteristics, and has gained the remarkable achievement applied in the field of image processing.In this thesis, we have extended PCNN's applications in image registration and image interpolation on the basis of deeply studying its theories and applications. Firstly, taking advantage of the translation, rotation and distortion invariance of PCNN, we propose a new image registration algorithm which considers the coordinate barycenter of neuron cluster as the characteristic point. And it has been applied to achieve functional magnet resonance medical image registration because of the stability of barycenter, the insensitivity to noise and geometric invariance. Experiment results of image registration have shown that the proposed algorithm can implement MR image registration fast, effectively and accurately. Compared to algorithms based on search strategy of optimization measure, it overcomes the local extremum problem and large computational complexity. Secondly, a new image interpolation algorithm, which can preserve the edges well, has been proposed using the firing cluster characteristic of PCNN and the propagation paths of pulse waves generated from the cluster process. The firing propagation paths are used to guide the interpolation of cluster inner. Then the nets function and rational fraction interpolations are introduced to deal with the mesh problem in the cluster inner and the peak problem in the cluster intervals caused by the interpolation based on firing propagation paths. Many simulation experiments of image interpolation have also indicated that our interpolation algorithm is effective and satisfying. Both of these new algorithms are the results of further useful exploration of PCNN's potential applications in image registration and image interpolation.