结合深度卷积网络与加速鲁棒特征配准的图像精准定位  被引量:8

Accurate image locating combining deep convolution network with SURF registering

在线阅读下载全文

作  者:罗家祥[1] 林畅赫 王加朋[1] 胡跃明[1] LUO Jia-xiang LIN Chang-he WANG Jia-peng HU Yue-ming(School of Automation Science and Engineering, South China Univ. of Tech. , Guangzhou 510640, China)

机构地区:[1]华南理工大学自动化科学与工程学院,广东广州510640

出  处:《光学精密工程》2017年第2期469-476,共8页Optics and Precision Engineering

基  金:国家科技重大专项(2014ZX02503-3);中央高校业务经费(2014z0033;2015ZM137)

摘  要:针对在大图像中定位小块区域图像的需求,本文提出一种结合深度卷积网络与加速鲁棒特征(SURF)配准的精准定位方法。将标准大区域图像分割成若干个小参考图像,利用深度卷积网络和类局部敏感哈希降维法提取参考图像集的特征并形成特征库;基于特征库,提出了先检索多个相似参考图像后再进行SURF精确配准的两阶段方法,实现目标小图像在标准大图像中的定位。针对电子工业过程中高密度柔性电路板(FPC)及精确末制导中的图像定位数据进行实验,实验结果表明,该方法避免了传统SURF算法大量的特征提取与配对过程,SURF特征提取数减少近90%;与直接根据图像特征进行配准的传统定位方法相比,在保证定位准确率的基础上,耗时可缩小一个数量级以上。For small-scale image locating in a large image,an accurate locating method combining deep convolution network with SURF registering was introduced.The large-scale image was divided into several small reference images,and the feature of such reference image set was extracted to form a feature library by combining the deep convolution network and Similar Local Sensitive Hashing(SLSH);on the basis of the feature library,a two-stage method that carrying out accurate SURF registering after retrieval of multiple similar reference images was put forward to achieve the locating of small target in a large image.Experiment was established on high density FPC and location data of accurate final guidance image,and the results indicate that the method avoids approximately 90%in the amount of feature extraction by comparing with traditional SURF locating methods,in which registering is directly carried out in accordance with image features.So the method can ensure the locatingprecision,meantime can lower the time consumption by more than one order of magnitudes.

关 键 词:深度卷积网络 图像检索 特征匹配 精准定位 

分 类 号:TH703[机械工程—仪器科学与技术]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象