Large Distortion Image Registration Based On ASIFT-LIKE Algorithm

In recent years,digital image technology has been continuously developed,and the research results on computer vision have been continuously applied in various fields,among which image registration,as a research hotspot,has been widely applied in many aspects,including image mosaic,change detection,image fusion,map drawing,etc.However,in image registration,due to the difference of imaging time and imaging angle,images will change in rotation,scaling,afine,etc.When these changes are large,it is the case that the geometric distortion of the image to be studied in this paper is large.Registration for such images will lead to inaccurate registration or registration failure caused by less feature points extracted for the same object and higher mismatch rate.Therefore,the specific work of this paper is as follows:Firstly,the data set is made and preprocessed.The data sets used in this paper are mainly 5 sets of public test data sets and 2 sets of self-made data sets for registering large distortion images.The adopted distorted images are classified according to whether the objects are single or not and the longitude and latitude changes of the images.In addition,considering that the existence of inevitable noise in actual situations will lead to a lot of outliers,the existing algorithm adopts the method of filtering before registration,which is difficult to deal with different kinds and levels of noise,and has low registration efficiency,long time and low robustness.Therefore,The adopted datasets are respectively added with different types and different levels of noise.Finally,a dataset containing 105 samples was obtained(35 with different levels of Gaussian noise,35 with different levels of salt and pepper noise,and 35 with different levels of speckle noise).Secondly,aiming at the problem of large distortion image registration and the robustness of the proposed algorithm to noise,an ASIFT-LIKE algorithm is proposed in this paper.In addition,aiming at the problem of uneven distribution of feature points extracted by the existing algorithm,which makes the image registration ineffective in some areas,a DPP-RANSAC method is proposed to eliminate mismatching,and the method proposed in this paper is used to extract and register the feature points of the data set.At the same time,the same data set is applied to general large distortion image registration algorithms(MSER algorithm,ASIFT algorithm.Harris-Affine algorithm,Hessian-Affine algorithm),compared according to experimental results.Finally,combining the evaluation criteria such as the number of registration points,the registration accuracy and the distribution of matching points,the effectiveness of ASIFT-LIKE proposed in this paper in the registration of large distortion images and its robustness to different types and levels of noise are proved,which ensures that enough feature points can be extracted while ensuring the extraction accuracy,and the distribution of feature points is more uniform.