Research On Classification And Recognition Of Scattering Spectra Based On Neural Network

Scattering spectrum can effectively show the information of material category,microstructure components and surface characteristics, and can be used to identify the target species and extract the properties of the target material. The complex three-dimentational structure and different components of many target material surfaces make the scattering spectra with overlapping and complicated nonlinear relationship between each other. And informative spectra are gotten which makes sample scattering spectra difficult to be identified in a very short period. Therefore, it needs further exploration for the scattering spectral scattering spectral classification and identification of this kind of materials. This paper mainly studies four kinds of different material samples are irradiated in the same light source, based on scattering spectra of different detection angles and the property of neural network self-learning, to identify and classify scattering spectra of different sample materials, and to achieve the purpose of classification and identification of the materials. The research contents are as follows:1. Established BRDF texture spectrum measurement system, scattering spectra of four kinds of sample material surface are measured under different detection and rotation angles and analyzed, and BRDF database of material spectrum is established based on the measurement result.2. Based on the characteristics of the artificial neural network and wavelet transform method, the noise of experimental measurement of scattering spectral data is smoothed and the signal noise ratio of the noise smoothed and the original noise is calculated. The scattering spectrum data of unknown material samples are identified and classified under350-750 nm band, and the accuracy rate of the result reaches to 95.54%. It is found that result of the signal noise ratio of noise smoothed is higher than original data, which shows that the noise can be effectively suppressed by wavelet. It helps to learn improves the learning speed and output time of RBF network and improves the accuracy rate for the classification and identification of material types.3. In order to improve the accuracy rate of identification and save the radial basis network learning time, scattering spectra under 350-750 nm band are selected and analyzed,and the data are normalized. By applying the theory of correlation, spectral characteristics of different wavelengths are analyzed. And the relevance of different material sample reflectivity under the same band is analyzed, the lower the correlation coefficient, the more differences in spectral lines and the more prominence of characteristics. Spectrum bands of385-506 nm are selected as characteristic spectrum band, which are classified and identified.And the result of accuracy rate is above 99%. It is found that selecting effective characteristic band helps to save text time and improves the recognition rate of materials.