摘要
针对基于深度学习的图像局部特征检测与描述模型参数大和内存资源消耗大的问题,提出一种结合注意力和多重特征融合的图像局部特征检测及描述算法。首先在不改变图片大小的条件下,使用基础残差块和膨胀卷积搭建基础骨干网络以获得图片的多尺度特征;然后在不增加模型复杂度的前提下,结合注意力机制从而获得更加优良的特征;最后通过改进的跳跃连接将低层语义信息与高级语义信息融合,更利于特征点的检测和描述。实验结果表明,在模型大小远小于同类方法模型的情况下,所提算法在HPatches数据集上表现优异,当阈值选取2~6时总体上的MMA值分别为0.57、0.71、0.78、0.81、0.83,与R2D2相比,分别提升了2.3%、2.4%、1.9%、1.4%、1.1%,提取的特征及描述更加鲁棒。
Aiming at the problem of numerous parameters and memory resource consumption for deep learning-based image local feature detection and description,a novel image feature detection and description algorithm combining attention scheme and multi-feature fusion is proposed. First,without changing the size of the picture,using the basic residual block and dilated convolution,we construct a basic backbone network to obtain the picture’s multi-scale features. Then,without increasing the complexity of the model,we obtain better feature by using the attention mechanism. Finally,the improved jumping connection combines low-level with the high-level semantic information,which is more conducive to the detection and description of feature points. Experimental results demonstrate that when the model size is much smaller than that of similar methods,the proposed algorithm performs well on the HPatches data set. When the threshold is selected from 2 to 6,the overall MMA value is 0.57,0.71,0.78,0.81,0.83,respectively.Compared with R2 D2,it has increased by 2.3%,2.4%,1.9%,1.4%,1.1% respectively,and the extracted features and descriptions are more robust.
作者
梁水波
刘紫燕
袁浩
孙昊堃
梁静
LIANG Shuibo;LIU Ziyan;YUAN Hao;SUN Haokun;LIANG Jing(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2021年第8期1075-1081,共7页
Chinese Journal of Sensors and Actuators
基金
贵州省科学技术基金资助项目(黔科合基础[2016]1054)
贵州省联合资金资助项目(黔科合LH字[2017]7226)
贵州省科技计划项目(黔科合基础[2017]1069)
贵州省科技计划项目(黔科合SY字[2011]3111)。
关键词
深度学习
特征检测
特征描述
注意力机制
特征融合
多任务学习
deep learning
feature detection
feature description
attention mechanism
feature fusion
multi-task learning
