摘要
针对由于拍摄视角不同,目标图像在水平或垂直方向发生拉长或压缩等仿射变换,进而无法正确识别的问题,本文设计了一种几何与仿生融合的特征提取方法 .首先,对传统的角点和直线检测进行改进,提出自适应Harris角点检测方法和去冗余的直线检测方法,并将角点数、直线数和面积比向量作为几何特征.然后,采用生物启发变换算法提取图像的仿生启发特征,该算法包括两个阶段,每个阶段均需执行方向边缘检测和局部空间频率检测.接着,将输入图像的两种特征向量分别与标准数据库中的特征向量进行Pearson相关距离计算,获得匹配得分.最后,在考虑不同数据库两种特征区分性强弱的基础上自适应确定权值,最高融合分数所对应的标签即为该图像的识别结果.实验结果表明,该方法能较好地提取图像的仿射不变特征,并且该方法在Alphanumeric,MPEG-7,GTSRB和MNIST数据库的识别准确率分别为92.2%,96%,90%和87.3%.
In view of the problems that the target image is elongated or compressed in the horizontal or vertical direc⁃tion due to different shooting angles,and the target image cannot be correctly recognized,this paper designs a feature extrac⁃tion method that combines geometry and bionic vision.Aiming at geometric feature extraction,this paper improves the tra⁃ditional corner and line detection algorithms,and proposes an adaptive Harris corner detection algorithm based on iterative threshold that draws on the similarity of regional pixels and a de-redundant line detection algorithm based on Hough trans⁃form.The number of corner points,the number of straight lines,and the area ratio feature vectors are taken as geometric features.Then,the bio-inspired transformation algorithm is used to extract the bionic visual features of the target image.The algorithm includes two stages.Each stage needs to perform directional edge detection and local spatial frequency detec⁃tion.Calculate the Pearson correlation distance between the geometric and bionic inspired features of the target image and the features in the standard database to obtain the matching scores.The weights are adaptively determined on the basis of considering the distinguishing strength of the two characteristics of different databases.And the label corresponding to the highest fusion score is the recognition result of the image.Experimental results show that the fusion method can well ex⁃tract the affine invariance features of affine images.The recognition accuracy of this method in Alphanumeric,MPEG-7,GTSRB and MNIST databases are 92.2%,96%,90%and 87.3%respectively.
作者
余伶俐
易倩
金鸣岳
周开军
YU Ling-li;YI Qian;JIN Ming-yue;ZHOU Kai-jun(School of Automation,Central South University,Changsha,Hunan 410083,China;School of Intelligent Engineering and Intelligent Manufacturing,Hunan University of Technology and Business,Changsha,Hunan 410205,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2023年第6期1607-1618,共12页
Acta Electronica Sinica
基金
国家自然科学基金(No.61976224,No.61976088)。
作者简介
余伶俐,女,1983年生,江西景德镇人.中南大学自动化学院教授.主要研究方向为移动机器人导航规划、视觉感知处理.E-mail:llyu@csu.edu.cn;易倩,女,1999年生,湖南长沙人.中南大学自动化学院硕士研究生.主要研究方向为图像处理、仿生视觉.E-mail:204612175@csu.edu.cn;通讯作者:金鸣岳,女,1996年生,河北邯郸人.中南大学自动化学院硕士研究生.主要研究方向为仿生视觉感知启发下的不变特征提取方法.E-mail:jinmingyue258@163.com;周开军,男,1979年生,湖南常德人.湖南工商大学智能工程与智能制造学院教授.主要研究方向为机器视觉、图像处理.E-mail:zkj@hutb.edu.cn。