Research And Application Of Classification For Ancient Porcelain Fragments Based On Microscopic Features

In the field of ancient porcelain identification,evaluating the age,the firing kiln and the category of the porcelain are still important parts of ceramic archaeology.During the process of protecting and researching the ancient relics,a large number of ancient porcelain fragments are often unearthed.These mixed cultural relic fragments belong to different types.They are from different artifacts and produced in different ages,which increases the difficulty of fragment classification.It also affects the speed of porcelain restoration.Therefore it is not convenient to manage and protect the cultural relics.This thesis focuses on the problem of ancient ceramic fragments automatic classification.Aiming for solving the problems of microscopic image preprocessing,surface scratch detection and microscopic image classification of ancient ceramic fragments,the theoretical analysis,data set production,method design and experimental comparison are used respectively.The main contributions of this thesis are:(1)Creating the microscopic images dataset of ancient porcelain fragments.Because of the lack of the complete release of microscopic image dataset in the field of microscopic research of ancient porcelain,the ancient porcelain microscopic image data integration for image classification should be figured out at first.An optical digital microscope camera was used to take microscopic images of the glaze layer of ancient porcelain fragments.According to the types of porcelain fragments,the microscopic images dataset of ancient porcelain fragments was classified which will provide data basis for the following work.(2)To solve the problems of unclear and uneven lighting of porcelain microscopic images,Gamma correction and CLAHE equalization algorithms are proposed based on the analysis and the research on image preprocessing methods.This method can enhance the contrast of ancient porcelain microscopic images and improve the clarity of ancient porcelain microscopic images.Because the scratch features(the disorder and depth of the scratches)on the surface of the porcelain are not obvious enough and difficult to distinguish.A scratch detection framework for the surface of the porcelain is proposed.Then the marks can be effectively detected and presented,which makes it easy for researchers to judge the degree of damage of the porcelain and confirm the generation age and other important attributes through the detected scratches.(3)Because of lacking the ancient porcelain fragments microscopic images,the method of transfer learning(fine-tuning)is used to accelerate the convergence of the deep network,which effectively reduces the computational cost during model training.To solve the problem of large memory and high consumption of neural network,the T-S model is proposed that trains the network pruning factor while fine-tuning the migration learning.Experiment results show that the T-S model proposed can reduce the training cost and shorten the classification time while ensuring the accuracy of fragment classification.