可见-近红外光谱的古陶瓷断代分类识别

doi:10.3964/j.issn.1000-0593(2019)03-0756-09

Abstract
References
Related Citation (15)

Download: PDF (5252 KB)
Export: BibTeX | EndNote (RIS)

Abstract Visible-near infrared spectroscopy based chronology classification of ancient ceramic method has been proposed to make the identification more objective and accurate. Yaozhou kiln exists in many dynasties and it has great similarity between different dynasties. Therefore, age identification of Yaozhou kiln faces great challenges. Taking Yaozhou kiln as the research object, some multi-spectral data of ancient ceramic from different dynasties are gotten from ultraviolet-visible near infrared spectroscopy analyzer. To avoid the first-order and second-order differential missing intermediate transition information, a fractional-order differential preprocessing method is proposed to suppress and eliminate the background information and noise from spectral data. The experimental results show that the classification accuracy of Yaozhou kiln in different dynasties is only 84.8% when the differential processing is not performed (0^th order), while the classification accuracy based on different fractional differentials is obviously higher than that of 0^th order. And the optimal order is 0.7. Then, a deep belief network based ancient ceramic classification method is proposed. First, stacked restricted Boltzmann machine (RBM) is employed to extract some high-level features during pre-training stage. The results show that the correlation coefficient between the features before RBM dimension reduction is 0.885 7, while the correlation coefficients after dimension reduction by the first and second RBM are 0.544 6 and 0.391 5 respectively, which means the redundancy is obviously cut back. Then some weight and bias values trained by RBM are used to initialize BP neural network. The whole deep belief network is fine-tuned by BP neural network to promote the initiative performance of network training and overcome local optimal limitation of the neural network due to the random initializing weight parameter. Experimentally, the optimal number of RBMs in depth belief network is 2, and the optimal number of RBM hidden layer units is 100. Meanwhile a dropout strategy is put forward to randomly ignore neurons of some hidden layers to reduce interdependence between features in the network training process and prevent over-fitting from some small data. When the ratio of Dropout is 0.45, the classification accuracy is highest. According to the method mentioned in this paper, the chronology classification accuracy in Yaozhou kiln is 93.5%, and accuracy of Yaozhou kiln in the Five Dynasties is highest, reaching 96.3%. Comparisons with some chronology classification methods highlight the superior performance of the developed method.

Key words：Visible-near infrared spectroscopy; Fractional order differential; Deep belief network; Dropout

Received: 2017-11-20 Accepted: 2018-04-25

ZTFLH:

TP391.7

Corresponding Authors: GUAN Ye-peng E-mail: ypguan@shu.edu.cn

Cite this article:

URL:

https://www.gpxygpfx.com/EN/10.3964/j.issn.1000-0593(2019)03-0756-09 OR https://www.gpxygpfx.com/EN/Y2019/V39/I03/756