基于深度学习的高速列车图像配准算法  

Image registration algorithm of bullet train based on deep learning for bullet train

作　　者：刘幸子 罗林[1] 张渝[1] 彭建平[1] 李金龙[1] LIU Xing-zi;LUO Lin;ZHANG Yu;PENG Jian-ping;LI Jin-long(School of Physical Science and Technology,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学物理科学与技术学院,成都610031

出　　处：《信息技术》2021年第7期26-30,共5页Information Technology

基　　金：中央高校基本科研业务基金(A0920502052001-241)。

摘　　要：列车车底部件检测通常采用对应区域比对策略,模板图像和当前图像因抖动造成差异,产生图像配准偏差直接影响检测结果。提出一种基于深度学习的图像快速配准算法,通过图像块的四角点单应性参数化,完成一幅图像至另一幅图像的映射,构成动车车底图像数据集,其中通过均方误差回归局部区域的角点的形变量,获得单应性矩阵,同时在预处理阶段,对图像进行直方图均衡化,放大特征,最后利用单应性矩阵对待配准图像进行仿射变换。为了验证模型的有效性,在动车车底图像上进行了测试,并对比几种经典模型,结果证明文中所提模型具有很好的性能。The detection of parts under bullet trains mostly adopt the corresponding area comparison strategy leading to difference between the template image and the current image caused by shaking,the image produced by which would affect the detection results.To solve the above problems,a 4-point homography parameterization which maps the four corners from one image into the second image,which constitute the image data set of train.The proposed model regards the affine transform as regression problem,and applies MSE(mean square error)to compute deformation between corners.Then homography matrix is obtained according to computed deformation.and the image is enhanced in the preprocessing stage.Finally,the registration image is transformed through the homography matrix.In order to verify the effectiveness,experiments are carried out on the constructed dataset,and several classical models are compared.The results prove that the proposed model is state-of-the-art.

关 键 词：光学 深度学习 图像配准 单应性矩阵 仿射变换 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]