| Image registration is not only a very important intermediate step in image fusion and 3D reconstruction,but is also receiving increasing attention in the field of distant perception and radiography.Therefore,the study of image registration methods is of great theoretical value as well as practical significance.Traditional alignment methods use the whole image for feature point matching,which introduces the interference of irrelevant regions and the lack of accuracy in feature point matching.This paper addresses the above problems by using an improved Faster RCNN neural network to extract regions of interest(ROI)from reference and floating images,avoiding the interference of irrelevant regions.The matched feature points are matched using a chunking-based bi-directional approach to reduce the number of mismatched feature point pairs.The main work in the study of the Faster RCNN based image alignment algorithm is as follows.Based on the speculative science and basic structure of the CNN elaborated,the ROI is extracted using an improved Faster RCNN.Firstly,the activation function is selected and the data image is trained on multiple scales,after that the number of anchor frames for mapping the original image is increased,and finally the online difficult sample learning method is introduced to achieve the improvement of Faster RCNN,and simulation experiments are conducted for the improved Faster RCNN method.Combining image registration basics and traditional theory,the improved arithmetic is used to suited characteristic spots.The absolutely range between the reference image and the floating image characteristic spots is first calculated,followed by coarse matching of characteristic spots by a two-way matching method,and finally the points with poor matching accuracy are filtered out using the chunking method and the RANSAC arithmetic to improve the precise of characteristic spots suited.The improved characteristic spots suited method is also simulated and experimented.Finally,a complete simulation experiment is carried out for the Faster RCNN based image registration arithmetic designed in this thesis.Firstly,the ROI suited characteristic spots is completed using the measure in this thesis,followed by the characteristic transformation experiments to obtain the final test outcome.The test outcome show that the RMSE of this measure is reduced by 0.425 on average compared with that of the SIFT measure,and therefore the Faster RCNN based image registration algorithm outperforms the traditional measure. |
