基于深度学习理论的红外和可见光图像配准  

Infrared and visible image registration based on deep learning theory

作　　者：刘洋[1] 杨晨[2] LIU Yang;YANG Chen(Dalian Neusoft University of Information,Dalian 116000,China;Dalian Jiaotong University,Dalian 116000,China)

机构地区：[1]大连东软信息学院,辽宁大连116000 [2]大连交通大学,辽宁大连116000

出　　处：《激光杂志》2020年第11期81-85,共5页Laser Journal

基　　金：国家自然科学基金项目(No.61881230684)。

摘　　要：以实现红外和可见光图像精准、快速配准为目标,提出基于深度学习理论的红外和可见光图像配准方法。通过参数传输使红外参数预训练模型与可见光模型的参数得到共享,赋予红外预训练模型目标检测能力,并利用采集到的红外图像数据展开模型预训练,得到深度学习的红外目标检测模型。利用仿射变换处理图像间的几何位置差异,运用Partial Hausdorff距离进行相似测度提取图像特征,最后采用人工免疫网络算法搜寻全局最优解,实现红外和可见光图像配准。实验结果表明:该方法可获取高精度的红外目标检测结果,检测精度高达92.8%,并且检测稳定性强,速率快,可有效修正红外和可见光图像间的灰度、旋转和平移等差异,具有较高的使用价值。In order to achieve accurate and fast registration of infrared and visible images,a registration method based on depth learning theory is proposed.Through parameter transmission,the parameters of infrared parameter pretraining model and the visible light model are shared.The target detection ability of infrared pre-training model is given.The infrared image data collected is used to expand the model pre-training,and the infrared target detection model of deep learning is obtained.The affine transform is used to process the geometric position difference among images,the Partial Hausdorff distance is used to extract image features,and the artificial immune network algorithm is used to search for global optimal solution to achieve infrared and visible image registration.The experimental show that this method can obtain high-precision infrared target detection results,the detection accuracy is as high as 92.8%,and the detection stability is strong,the speed is fast,it can effectively correct the gray level,rotation and translation differences among infrared and visible light images,and it has a high use value.

关 键 词：深度学习 红外 可见光 图像配准 人工免疫 

分 类 号：TN391.4[电子电信—物理电子学]