Research On Attention-HardNet Feature Matching Algorithm In Sub-window Scale Space

Feature-based image matching algorithm is a core task in the field of computer vision.How to accurately and effectively match image features is an important topic to promote the deep learning of image and visual computing.Traditional feature matching algorithms suffer from serious loss of image corner information,long time-consuming,and severely affect the stability and accuracy of subsequent feature extraction,feature description,and feature matching steps.Feature extraction generally performs floating-point calculations,resulting in low efficiency in extracting feature points and poor compatibility of feature descriptor algorithms.Aiming at the above problems,this paper proposes an Attention-Hard Net feature matching algorithm in sub-window scale space.First of all,the scale space is constructed by the Sub-window box filter.The Sub-window box filter performs adaptive sub-window filtering on the image content,and uses the box filter combined with the sub-window regression to fully retain the corners without consuming a large amount of calculation.Constructing scale space efficiency,retaining more detailed information for feature extraction and feature description;secondly,using FAST algorithm for scale space feature point extraction,circular non-maximum suppression algorithm optimizes it,and setting different suppression radii and point response strengths Screen the feature points with strong stability;again,add the SENet attention mechanism to Hard Net to form Attention-Hard Net,and obtain the importance of each channel through learning,so that the network can extract more robust 128-dimensional floating-point features Descriptor;Finally,the L2 distance is used to measure the similarity of different descriptors to complete the image feature point matching.The performance test results of anti-scale,compression,and illumination of the matching algorithm on the Oxford data set show that the proposed algorithm can protect the edge and corner information of the scale space,improve the speed of feature point extraction,and increase the uniqueness of the descriptor and the reliability of the feature matching algorithm.Compared with deep learning methods such as L2 net and Hard Net,this algorithm improves the matching accuracy by about 3% and the speed by about 10%.There are 29 figures,4 tables and 53 references in this paper.