摘要
提出一种新的以边界不变矩作为识别特征,运用BP网络识别扩展目标的方法。首次通过详细的理论证明和实验分析,揭示了离散边界不变矩不再具有严格的比例不变性,而位移和旋转不变性保持相对稳定,并对该不变矩作为识别特征的误差进行了深入分析,给出了正确计算边界不变矩的途径。在此基础上,以该边界不变矩作为识别特征,输入BP网络,采用合理的网络结构,实现对发生位移、旋转和尺度变化的扩展目标的识别。边界不变特征的引入,减少了数据运算量,实验结果表明,识别率达到95.9%。
In view of recognition efficiency and movement complexity of extended targets, a new extended target recognition method with BP neural network is proposed, in which moment invariants based on target boundary are utilized as recognition features. The RST （Rotation, Scale and Translation） invariance of such features in digital condition is proved and analyzed for the first time. The analysis results say that their scaling invariance is lost, while translation and rotation invariance are almost maintained. And also the errors of these invariants during target recognition are analyzed and the right way to get these recognition features is given. Using these features as inputs, through BP neural network with reasonable structure, the extended target is recognized correctly. With this method, computation is reduced by introducing invariants based on boundary. Experimental results indicate the recognition ratio reaches 95.9%.
出处
《光电工程》
EI
CAS
CSCD
北大核心
2005年第8期1-5,共5页
Opto-Electronic Engineering
关键词
目标识别
扩展目标
BP神经网络
特征提取
Target recognition
Extended target
BP neural network
Feature extraction
