摘要
为获取高质量的粒子图像,提出一种双分支残差卷积神经网络DBRNet用于粒子图像测速(PIV)技术中的粒子图像增强。首先,设计一种由残差块组成的双分支卷积神经网络对输入的粒子图像对进行特征提取,同时,采用编码-解码器对粒子图像对的特征信息进行有效融合。其次,自主生成具有挑战性的图像增强数据集训练模型参数,其中,包含不同浓度的高斯噪声、光强噪声及多种真实的干扰背景,以充分模拟真实流体场景。结果表明,本文方法能够有效处理合成图像和真实图像中的噪声干扰,实现图像增强;利用速度场估计算法处理经本文方法增强后的粒子图像对可以得到更高精度的速度场。
A dual-branch residual convolutional neural network was proposed for image enhancement in PIV velocity technique to obtain high-quality particle images.Firstly,a dualbranch convolutional neural network composed of residual blocks was designed to extract features from the input particle image pairs,while a coding-decoder was used to effectively fuse the feature information of the particle image pairs.Secondly,a challenging image enhancement dataset was autonomously generated to train model parameters,including Gaussian noises of different concentrations,light intensity noise and various real interference backgrounds,thereby fully simulating real fluid scenes.Results show that the proposed method can effectively deal with noise interference in both synthesized and real images,achieving image enhancement.Meanwhile,higher precision velocity fields can be obtained by using velocity field estimation algorithm to process the particle image pairs enhanced by the proposed method in this paper.
作者
张志浩
于长东
刘百胜
范毅伟
ZHANG Zhihao;YU Changdong;LIU Baisheng;FAN Yiwei(College of Artificial Intelligence,Dalian Maritime University,Dalian 116026,China;College of Software,Liaoning Technical University,Huludao 125000,China;College of Shipbuilding Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《大连海事大学学报》
CSCD
北大核心
2024年第4期100-109,共10页
Journal of Dalian Maritime University
基金
国家自然科学基金青年基金资助项目(52401362)
中央高校基本科研业务费专项资金资助项目(3132024224)。
作者简介
张志浩(2000-),男,研究生;于长东(1996-),男,博士,讲师,E-mail:ycd@dlmu.edu.cn。
