Classification Of Ancient Ceramic Relic Fragments Based On Learning Optimization And Information Fusion

In the process of Computer-aided ceramic restoration,that is a NP hard problem to splicing a large number of fragments.Efficient intelligent classification of fragments can reduce the size of the exhaustive enumeration of debris by dividing the subset of fragments,decomposition and moving forward the difficulty of the splicing NP hard problem,to form an effective support for the computer aided repair of broken cultural relics.Through theoretical analysis,method design and experimental comparison,this paper focuses on the intelligent classification of ceramic fragments from following key aspects:feature extraction methods of ceramic artifacts,process optimization of active learning classification,quality codebook generation of bag-of-word model,classifier ensemble method based on decision fusion,furthermore,the feature representation and classification of ceramic artifacts by fusion of color emotion applied.The research work has been supported by the National Natural Science Foundation of China,and a lot of contributions are made in this paper:(1)A new method to describe and compare the visual features of ceramic artifacts is proposed.A robust feature extraction method based on shape contour sampling is proposed for porcelain artifacts,and the similarity measure method of the decorative features based on quantum genetic algorithm is presented;We get porcelain glaze texture features by based on the sparse representation theory,and extention of the classical gray level co-ccurrence matrix method,which method integrates color features and texture features effectively.For the Pottery relics,through the improvement of the key points in the classical 2DSift method,such as the location of key points and the determination of direction,the extraction of significant geometric features of 3D point cloud is realized based on cluster analysis.Through the experiment of the porcelain image and terracotta tiles pieces,the proposed method of feature representation and similarity measurement is more in line with the characteristics of cultural relic fragments,which can effectively realize the visual feature representation and comparison of ceramic cultural relic fragments.(2)To solve the problem of insufficient labeled historical relic samples,an active learning optimization classification method based on representative sample selection is proposed.As to the rare and precious cultural relics,among the many categories,there are few labeled samples of the same category,which can not effectively support the full training in the training process.Based on the idea of active learning,the study of the interactive feedback mechanism to achieve the goal of effective selection of samples,an active learning optimal classification method of combining sample uncertainty measurement and redundancy elimination is proposed in this paper by using the kernel fuzzy clustering to alleviate the problem of the insufficient initial training samples and learning in the training process.The experimental results show that the proposed method can effectively improve the efficiency of correlation feedback in the active learning process,and further realize the training generation of efficient classifier.(3)The research has optimized the method of establishing the code book of Bag-of-Words model.Because of the complex codebook construction process and inefficient analysis of the redundant relations between the generated visual words,the existing method may have a negative effect on the final results of the classification.In order to solve this problem,this paper improves the process of clustering codebook generation and optimizes the initial codebook according to the dependency relation between words to obtain the visual lexical codebook with high distinguishing ability finally,which brings good results in the classification experiments of debris images.In the experiment of ceramic debris image classification,a highly separating capacity codebook was obtained by this method,and the classification accuracy was improved by 20%compared with the traditional word bag model classification method(4)A method of classifier ensemble based on decision fusion is proposed.Ceramic artifacts have multiple heterogeneous features.Because of the classification system based on single feature and single classifier is limited by the objective factors such as insufficient information utilization and the defects of classification methods,so,which is less practical.In order to solve the problem,the idea of the difference measure of sub classifiers is introduced to the harmony search algorithm,and an adaptive global optimization and harmony algorithm(AGOHS)is proposed firstly by measuring the diversity of the harmony database,and then,based on the idea of decision fusion and the comprehensive measurement of classification accuracy and diversity,the ensemble of classifier based on AGOHS is realized,and in this way,an efficient classification system is constructed by adding information between sub classifiers.The practical effect of comprehensive application of ceramic fragments is improved.(5)A new method for classification of ceramic fragments based on color emotion feature fusion is proposed.The glaze color and decoration are the most important attributes of Chinese cultural relics,and the artistic styles of ceramic artifacts are closely related with the major characteristics of the color and decoration.The image contains a wealth of emotional information,and the domain color of image can be extracted by fuzzy clustering based on spatial constraint at first,and then calculate the color emotion feature based on PAD model to construct ceramic glaze sparse dictionary,finally,it is applied to the classification of ceramic fragments and good experimental results are obtained.This is a good exploration of the expression and application of color emotion in the characteristics of ceramics,and laid the digital technical foundation of the further research on ceramic art style.