结合深度卷积网络与加速鲁棒特征配准的图像精准定位

发布时间：2018-06-05 18:45

本文选题：深度卷积网络 + 图像检索　；参考：《光学精密工程》2017年02期

【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。
[Abstract]:Aiming at the need of locating small area images in large images, this paper presents an accurate localization method combining depth convolution network and accelerated robust feature surf registration. The standard large area image is divided into several small reference images, and the feature of reference image set is extracted by using depth convolution network and local sensitive hash reduction method. In this paper, a two-stage method of retrieving several similar reference images before accurate SURF registration is proposed to locate small target images in standard large images. Experiments are carried out on high density flexible circuit board (FPC) and image location data in precision terminal guidance in electronic industry. The experimental results show that, This method avoids the reduction of the number of features extracted by the traditional SURF algorithm by nearly 90 points, compared with the traditional localization method, which is directly based on the image feature registration, on the basis of ensuring the accuracy of the location. The time taken can be reduced by more than one order of magnitude.
【作者单位】：华南理工大学自动化科学与工程学院;
【基金】：国家科技重大专项(2014ZX02503-3) 中央高校业务经费(2014z0033,2015ZM137)
【分类号】：TP391.41;TP183

【参考文献】