Adaptive “3R” De-Noising Algorithm Based on Near Infrared Bi-Spectrum
ZHAO Xiao-yu1，2, FANG Yi-ming1, TAN Feng2, WANG Zhi-gang3, TONG Liang4
1. Institute of Electrical Engineering, Yanshan University, Qinhuangdao 066004, China 2. College of Information Technology, Heilongjiang Bayi Agricultural University, Daqing 163319, China 3. College of Life Science and Forestry, Qiqihar University, Qiqihar 161006, China 4. Communication and Electronic Engineering Institute, Qiqihar University, Qiqihar 161006, China
Abstract：Adaptive de-noising algorithm is proposed based on transmission spectrum and absorption spectrum of near infrared. Near infrared transmission spectrum and absorption spectrum collected synchronously are decomposed into intrinsic mode functions by ensemble empirical mode decomposition; the intrinsic mode function is a single frequency component. Correlations between intrinsic mode functions and transmission spectrum, absorption spectrum were calculated, and the correlation between intrinsic mode functions of transmission spectrum and absorption spectrum was also computed. The results show that the intrinsic mode function with minimum correlation coefficient should be noise component. The self-correlation of this intrinsic mode function was analyzed to judge whether the intrinsic mode function is noise. IF the self-correlation is very large at the midpoint and is zero or very small at the other point of the spectrum, then the intrinsic mode function is noise component for judgment, based on which “3R” algorithm is named to judge whether the intrinsic mode function is noise component. Removing noise component, constructing spectral signal and circulating the previous decomposition was conducted, and the noise reduction process was ended until it did not meet the “3R” rule. To do experiment on the simulated spectrum with noise, the effect of de-noising with “3R” algorithm is better than EMD and EEMD low pass filter, and it is not so good as wavelet decomposition. In the real spectrum testing, the model was established between spectra treated by above methods with chlorophyll on three layers. BP neural network, and the model de-noised by “3R” method has the biggest correlation coefficient and prediction coefficient, but the smallest correction standard error and prediction standard error. “3R” method’s effects on the peak position and peak intensity of spectrum are the smallest among the four kinds of de-noising methods. Experiments show that the “3R” algorithm based on bi-spectrum can be used for near infrared spectra de-nosing without presetting the number of iterations, there is no need to consider layers of decomposition, also no need of basis function, and the adaptability is very strong.
