摘要
大数据时代,图像是重要的信息传递媒介,但图像质量退化将影响信息识别.针对各种类型的图像退化问题,提出一种融合长短期记忆（LSTM）的深度卷积神经网络（DCNN）的带记忆分类方法,识别退化图像模糊类型及其参数,根据准确的模糊类型及模糊核进行图像去模糊.首先改进DCNN卷积模型,调节卷积运算步长算子,加快图像卷积收敛速度;引入串行LSTM网络,将训练过的图像微元进行记忆,提高识别速度和准确性;通过BP网络输出模糊类型及其参数,再进行图像反卷积去模糊.实验表明能识别出3种主要模糊类型并识别率在90%以上,模糊参数误差在一个像素内,能复原出清晰的图像;最后将算法应用到实际高速铁路轨道缺陷检测系统中,对质量较差的图像进行模糊识别及去模糊,提高图像识别率.
Images have become one of the most important information carriers in the modern society,but image degradation affects object recognition. Aiming at improving the image quality,the blur types and parameters of degraded images were identified by fusing the DCNN（ Deep Convolutional Neural Network） and the LSTM（ Long Short-Term Memory） to improve performance with memory function. Firstly,the DCNN structure was designed,and the step length of the convolution operator was adjusted,which accelerates the convergence of image convolution. The serial LSTM was fused to memorize the trained features,which improves the recognition efficiency and accuracy. Then use the BP as the output net. Experiments show that the three main blur types can be identified and the accuracy is above 90%,the blur parameter errors are within one pixel and a clear image can be acquired. Finally,the method was applied to the railway-defects detection system in terms of blur identification and deblurring for degraded image so as to improve the image recognition rate.
作者
黄绿娥
吴禄慎
陈华伟
HUANG Lye;WU Lushen;CHEN Huawei(School of Mechanical and Electrical Engineering,Nanchang University,Nanchang 330031,China;College of Applied Science,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《应用基础与工程科学学报》
EI
CSCD
北大核心
2018年第5期1092-1100,共9页
Journal of Basic Science and Engineering
基金
国家自然科学基金项目(51065021,51365037)
江西省科技厅项目(20161BAB216128)
