TP391
国家自然科学基金资助项目（51405044）；机械传动国家重点实验室开放课题（SKLMT-KFKT-2017）。
数字图像相关方法中，位移场测量的误差大小与算法的迭代次数通常成反比，要获得较低的误差，必须增加迭代次数，从而增加了计算量；而非迭代的方法误差相对较大。为解决这一问题，提出了一种基于BP神经网络的误差补偿方法。选择基于非迭代光流法的位移场测量方法为算法模型，详细分析了该算法本身存在的截断误差，以模拟散斑图的位移测量值及其误差为数据集，用训练好的神经网络误差预测模型对测量结果进行补偿。实验验证结果表明，补偿后的位移测量误差相较原来总体下降了50%左右，测量误差的统计分布也显著下降。
The error of digital image correlation (DIC) displacement field measurement is always conflicted with the iteration times of algorithm. To reduce the calculation error, the number of iterations has to be increased, which will result in a heavy computing burden. However, the error of the non-iteration method is usually high. To solve the problem, a BP neural network-based error compensation method is proposed. The non-iteration optical flow algorithm is selected as analytical example and its error is also analyzed. The displacement measurement of a simulated speckle image and its error are used as training data. The displacement measurement result is compensated by the predicted value. The compensation experiment is carried out and it shows that the error after compensation drops by 50% and the histogram of the error is also reduced.
蒋中宁,罗远新,王勇勤,郭平.数字图像相关位移场测量的误差补偿[J].重庆大学学报,2020,43(12):59-67.
复制
