The main purpose in various methods of image registration is to find the transformation parameters for accurate mapping an image onto another image coordinates. In medical sciences creating a precise mapping between medical images data is very important in application More
The main purpose in various methods of image registration is to find the transformation parameters for accurate mapping an image onto another image coordinates. In medical sciences creating a precise mapping between medical images data is very important in application such as diagnosis and treatment. Accordingly, several approaches have been proposed for image registration. The compression of results and performance between different image registration algorithms was the main motivation for this research to design and implement a new hybrid algorithm so that provide high accuracy in multimodal image registration. Automating the image registration process by using machine learning approach is the innovation of this method compared to previous ones.
To this end, the proposed method which is named multi resolution learning is composed of multi resolution decomposition and a hierarchical neural network which it learn the transformation parameters by using global properties of the image and uses learned transformation parameter for image registration. The proposed method is implemented and tested on the medical images of Vanderbilt university database. Experiment result show acceptable accuracy for the proposed method compared with other methods.
Manuscript profile
Cellular learning automaton is an intelligent model as a composition of cellular automaton and learning automaton. In this study, an extended algorithm of cellular learning automata is proposed based on transfer learning as the TL-CLA algorithm. In this algorithm, trans More
Cellular learning automaton is an intelligent model as a composition of cellular automaton and learning automaton. In this study, an extended algorithm of cellular learning automata is proposed based on transfer learning as the TL-CLA algorithm. In this algorithm, transfer learning is used as an approach for computation deduction and minimizing the learning cycle. The proposed algorithm is an extended model based on merit function and attitude vector for transfer learning. In the TL-CLA algorithm, the value of the merit function is computed based on the local environment, and the value of the attitude vector is calculated based on the global environment. When these two measures get the threshold values, the transfer of action probabilities causes the transfer learning from the source CLA to the destination CLA. The experimental results show that the proposed TL-CLA model leads to increment the convergence accuracy as 2.7% and 2.2% in two actions and multi-action standard environments, respectively. The improvements in convergence rate are also 8% and 2% in these two environments. The TL-CLA could be applied in knowledge transfer from learning one task to learning another similar task
Manuscript profile
