Research On Dimensional Reduction Technology Of Semi-supervised Features And Application Of Classification Of Terracotta Warriors

The Terracota Warriors have been affected by natural and human factors all year round,and most of the excavations have been broken into piles.The cost of manual fragment classification restoration is very expensive,so the number of labeled fragments after manual marking is very small,and the number of unmarked fragments is large,which brings great difficulties to the classification of computer-aided cultural relic fragments.Traditional machine learning can only use unlabeled data for unsupervised learning or only labeled data for supervised learning,and cannot use both kinds of data at the same time.In order to obtain better learning accuracy and make full use of existing data,semi-supervised learning has become a hot research topic.The continuous increase of data not only leads to the increase of data dimension,but also affects the running speed and performance of related algorithms.Feature dimension reduction is the best way to solve this problem.Feature dimensionality reduction aims to retain the feature subset that can represent all information of the whole data set,and delete irrelevant features and redundant features.The results of dimensionality reduction directly affect the classification results of data.At present,the most commonly used methods of dimensionality reduction are feature extraction and feature selection.However,in practice,using only labeled data or unlabeled data can not represent all the information of the original data,which will lead to the feature after dimensionality reduction can not train the model well,the current research hot spots should combine all labeled and unlabeled data,Using the category of labeled data and the large amount of information hidden by unlabeled data to reduce the dimension of the data.Therefore,in this thesis,semi-supervised learning is combined with two methods of feature dimensionality reduction to carry out the pre-processing work before the classification of terracotta fragments.The main contents are as follows:(1)A semi-supervised feature extraction algorithm based on convolutional neural network is proposed by combining semi-supervised clustering technology and feature extraction method of convolutional neural network.On the one hand,semi-supervised clustering is used to assign labels to unlabeled data to train the convolutional neural network.On the other hand,the labeled data is used to adjust the network,and the final trained network is used for feature extraction and extraction effect evaluation of the image.This paper conducts experiments on CIFAR10 and The Terracota Warriors data sets,and compares the experimental results with other semi-supervised feature extraction algorithms.The results show that the feature extraction algorithm in this paper has better feature extraction capability than other semi-supervised feature extraction models,and can effectively extract useful features of images to achieve dimension-reduction effect.(2)Aiming at the problem that the current feature selection algorithm uses a single measurement method to select features,combining with filtering feature selection technology,a semi-supervised feature selection algorithm based on multiple measures is proposed to deal with the irrelevant and redundant features that still exist in the extracted feature set.In the algorithm,distance measurement and information theory measurement are used to select the relevant features,delete the irrelevant features and weak correlation features,and use the approximate markov blanket algorithm to delete the redundant features in the relevant feature set and retain the strong correlation features.After application of the algorithm in the feature extraction of characteristics,with the other a semi-supervised feature selection algorithm and feature extraction after the classification results of the comparison,the experimental results show that the feature selection methods are more effective than other models of choice,at the same time,the feature selection algorithm can significantly improve the characteristics of the discriminant ability,can effectively select features strong correlations of concentration,improve the classification accuracy of the models.(3)In order to measure the effectiveness of dimensionality reduction features,a semi-supervised debris classification system for The Terracota Warriors is designed and implemented.The feature extraction and feature selection methods involved in this paper are integrated into the system as the processing steps of the classification of terracotta fragments to realize the feature processing and classification functions of the image of fragments.