Research On The Technology Of Image Registration Based On Feature Points And Tow-point Entropy

There are some geometric transformation relationship between several images which come from different imaging equipment and different imaging conditions toward the same scene. If we want to comprehensively utilize image information from different conditions, it is necessary to match and align these images firstly, so they form a "synthesis" that can reflect more complete information about the scene itself. The application field of image registration is very wide, but at present in the image registration process there are some problems to be solved and these problems include: looking for more robust similarity measure, finding more efficient optimization strategies, seeking to avoid falling into local optimum method, finding ways to improve registration accuracy and speed and improving the real-time processing of high quality image and mass data. In order to solve these questions:to avoid falling into local optimum, find a better similarity measure, improve the precision and accuracy of image registration, this paper has done many study work to the registration process and method which mainly contain the following two aspects:Firstly, aiming at the shortcomings of image registration algorithm based common feature, this paper proposes an image registration method based on improved CPSO optimization. During RANSAC remove the false matching, it needs multiple randomly selected from a fixed number of matching points to determine the parameters and then quite a few points to verify this parameter, the computational complexity is very large. Therefore, this paper puts forward a simplified extended linear invariant moment method, which uses invariant moment’s rotation invariance to upgrade operation efficiency deducing the image error matching. Basely, image registration based on feature points uses Euclidean distance as the similarity measure, under which vectors are mapped to the corresponding scalar expression similarity which has not discriminative enough, it will result fauls matching, this paper solve the question by introducing probability strategy of chaotic particle swarm optimization algorithm has been improved. The experimental results show that the improved CPSO algorithm can achieve the image registration better and faster.Secondly, mutual information can be used to analyze the correlation of two images in image registration, and when the pixel coordinates of the two images are one towards one the correlation is the strongest. Namely when two images are most relevant the maximum mutual information value is most high and they are full registration. This paper improved mutual information-based image registration algorithm by using two-point entropy as a similarity measure, optimized by CPSO and based on TPE, it can solve the question that impact from original image’s gray level is relatively large in the mutual information registration method. The results of simulation experiments show that: compared to the mutual information, TPE and the improved Powell algorithm, the image registration algorithm based on TPE optimized by CPSO has better registration accuracy..