摘要
Abstract
The traditional feature point detection algorithm is sensitive to illumination changes and visual angle changes,which leads to the decline of feature positioning accuracy and repeatability,and restricts the accuracy and robustness of feature matching.In view of this,the paper proposes a feature point detection algorithm combining residual connections and multi-scale attention mechanism,and the algorithm is based on SuperPoint network.Firstly,residual connections are introduced into the feature point encoder to mitigate feature degradation in deep networks,preserving low-level feature integrity for enhanced extraction precision and repeatability.Secondly,a multi-scale attention mechanism is built and the feature information of different scales is taken into account to make the network capture the details and context in the image more comprehensively,so as to enhance the representation for features.Comparative experiments on the HPatches dataset demonstrate the effectiveness of the proposed method.Under illumination changes,the proposed algorithm achieves 5.9%higher feature repeatability than baseline SuperPoint,with homography estimation accuracy improvements of 29.3%,0.76%,and 0.73%at pixel error thresholds(ε)of 1,3,and 5,respectively.For viewpoint variations,repeatability increases by 11.2%,accompanied by increase of homography estimation accuracy of 9.1%(ε=1)and 3.9%(ε=3).Matching experiments reveal approximately 1%higher overall matching scores and significantly increased correspondence counts under both scenarios.These results validate the superior robustness and generalizability of the proposed method in the feature detection tasks.关键词
特征点检测/SuperPoint/残差结构/多尺度注意力机制/鲁棒性/HPatches数据集Key words
feature point detection/SuperPoint/residual connection/multi-scale attention mechanism/robustness/HPatches dataset分类
信息技术与安全科学
