Dropout Deep Belief Network Based Chinese Ancient Ceramic Non-Destructive Identification.

Area of Science:

• Archaeometry
• Materials Science
• Computer Science

Background:

• Accurate identification of ancient ceramics is crucial for cultural heritage studies.
• Non-destructive methods are preferred to preserve artifacts.
• Multi-spectral data offers rich information for material analysis.

Purpose of the Study:

• To develop a non-destructive identification method for ancient ceramics using multi-spectral data.
• To enhance spectral detail extraction and improve recognition performance.
• To address challenges like small sample sizes and overfitting in spectral data analysis.

Main Methods:

• Developed a dropout deep belief network model for ceramic identification.
• Applied a fractional differential algorithm for spectral data pre-processing and enhancement.
• Utilized unsupervised Restricted Boltzmann Machines (RBM) for feature extraction and pre-training.
• Employed back propagation (BP) neural network for fine-tuning the deep belief network, initializing weights from RBM.

Main Results:

• The fractional differential algorithm effectively enhanced spectral details.
• The dropout deep belief network demonstrated robust feature extraction.
• The proposed method achieved excellent recognition performance for ancient ceramics.
• The fine-tuning process using BP network overcame local optima and improved training stability.

Conclusions:

• The developed dropout deep belief network combined with fractional differential pre-processing offers a highly effective non-destructive method for ancient ceramic identification.
• This approach successfully mitigates overfitting issues common with small spectral datasets.
• The method shows significant potential for applications in cultural heritage and archaeological research.