摘要
针对光晕导致的光斑图像边缘模糊的特点,采用Niblack局部阈值分割得到光斑目标区域,并提取光斑的几何特征;以Niblack分割得到的图像边缘对原始光斑图像进行裁剪,得到去除光晕影响的光斑目标灰度图像,在此基础上提取该区域光斑图像亮度,结合光斑几何特征构造6维特征矩阵。分别采用BP神经网络、线性局部切空间排列LLTSA-BP网络、局部保持投影LPP-BP模型对烧蚀功率进行识别;进一步采用极限学习机(ELM-Extreme Learning Machine)、LLTSA-ELM和LPP-ELM降维模型,基于降维后的特征矩阵进行烧蚀功率分类。对比研究发现BP神经网络在对6维特征矩阵分类时收敛时间比ELM分类模型短,所需隐含层神经元个数少。而流形学习-ELM模型则在对降维之后的数据分类时表现较优,所需时间远远小于BP神经网络模型的处理时间,其中LPP-ELM模型对光斑的分类效果最优。
Aiming at the image feature of blurred edge caused by halo,the spot target area is obtained by Niblack local threshold segmentation,and then the geometric features of spot image are extracted;using image edges segmented by Niblack to clip original spot image,and grayscale image of spot target that have been removed from the halo is obtained,on this basis,the spot area brightness is extract,and the 6-dimensional feature matrix combined with the geometric characteristics of spot is constructed.The ablation power is identified by BP neural network,Linear locally tangent space arrangement(LLTSA-BP)and Local preserving projection(LPP-BP)models respectively;further,ablation power is identified based on characteristic matrix after dimension reduction by ELM(Extreme Learning Machine),Linear locally tangent space arrangement(LLTSA-ELM)and Local preserving projection(LPP-ELM)dimensionality reduction models respectively.Through comparative studies,we found that convergence time classifying 6-dimensional feature matrices of BP neural network is shorter than that of ELM classification model,and the number of required hidden layer neurons is less.However,kinds of manifold learning-ELM models,be used to classify data after dimensionality reduction,performs better,which takes much less time than the BP neural network do,and among them,the LPP-ELM model has the best classification effect.
作者
王福斌
刘洋
霍晓彤
李占贤
潘兴辰
WANG Fu-bin;LIU Yang;HUO Xiao-tong;LI Zhan-xian;PAN Xing-chen(School of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China;Hebei Robot Industrial Technology Research Institute,Tangshan 063200,China;Petro China Beijing Gas Pipeline Corporation Limited,Beijing 100101,China)
出处
《激光与红外》
CAS
CSCD
北大核心
2020年第1期117-123,共7页
Laser & Infrared
基金
国家自然科学基金项目(No.71601039)资助。
关键词
飞秒激光
极限学习机
流形学习
图像分类
femtosecond laser
limit learning machine
manifold learning
image classification
