Research On Image Feature Matching Algorithm Of Lunar Surface

Image processing technology is its core technology in computer vision tasks.In image processing technology,image feature detection and matching are the key technical links,which play a key role in many fields,such as target recognition,image fusion,and image splicing,3D reconstruction,target tracking,visual positioning,etc.The rapid progress of image processing technology has prompted researchers to transition objects from general natural objects to more complex and unknown cosmic objects.With the success of the moon landing and the development of satellite technology,the feature detection and positioning of the lunar surface have aroused high interest of researchers.This article takes the lunar surface image as the research object to study the image feature detection and matching algorithm.The main innovation of the paper is that it proposes a moon surface feature matching algorithm based on deep learning,which improves its matching accuracy in different environments.The specific research work is as follows:(1)Based on the research of traditional image feature matching theory,the traditional SIFT(scale-invariant feature transform)feature matching algorithm is used for feature matching of lunar surface images,which is mainly divided into feature point detection,feature point description and feature point matching.There are three parts,among which the detection of feature points is the core part of the algorithm,which is divided into scale space extreme point detection,feature point precise positioning,and feature point direction determination.Experiments verify that the SIFT matching algorithm has good stability in terms of scale changes,illumination changes,and rotation changes,but to a certain extent,problems such as mismatching and repeated matching will occur.(2)Applying deep learning to the field of image matching,a self-supervised learning feature point matching algorithm is proposed.The algorithm is based on the Super Point network and uses self-rendering to generate image training data.Combining isomorphic adaptation technology to increase the diversity of training data sets,Better improve traditional algorithms to achieve the disadvantages of complexity,low precision,and poor real-time performance.Experiments show that feature matching based on learning is better in a complex and changeable environment,and the matching effect is better for changes in the environment.(3)Combined with the transfer learning(Distillation)method,an image feature point matching algorithm based on multi-task learning is proposed,a six-degree-of-freedom visual positioning detection method is introduced,a large-scale local hierarchization is used and a lightweight network is used for feature point detection And description,the use of local and global descriptors to synergize the description of feature points to achieve a more accurate and effective feature point description,combined with the feature point detection sub-network of the Super Point network,and the multi-task distillation method of the teacher network for multi-tasking Synchronous learning to achieve more accurate and effective feature matching.Experiments and data analysis through the lunar surface image prove that it has more significant effects on scale transformation,rotation change,and illumination transformation.