Edge Intelligence in Internet of Things using Machine Learning 2022View this Special Issue
Semantic Segmentation-Based Adaptive Mining Algorithm for Ceramic Cultural Resource Data
Ceramic culture as Chinese culture has a long history and is the Chinese people’s spiritual home. The ceramic culture resource base contains a large number of images, videos, and other resources, so the collection of image category of data mining, as well as the finishing processing, is very critical. Thousands of years of ceramic data accumulation and impatient information demand have created a new point of contradiction for ceramic cultural resources. Therefore, in order to address this issue, we carried out a research study based on the concept of big data mining of ceramic cultural resource data, which is based on data fusion and feature extraction methods. We also considered semantic segmentation processing methods, which are used for data information management, scheduling, identification, collection, statistics, and aggregation of heterogeneous ceramic cultural big data. Further, a fine mining method for ceramic culture big data based on semantic segmentation is also proposed. As a result, the distributed storage of information and detection capability of ceramic cultural resource data are improved. The experimental results reveal that the proposed method performed better than the earlier approaches.
Ceramics occupies an important place in Chinese culture, and Chinese porcelain-making technology was a world leader throughout its long history until the late Qing Dynasty. Ceramic culture as Chinese culture has a deep heritage and is the spiritual home of the Chinese people. Ceramic culture is based on an immaterial form of exquisite skills, a distinct style of thinking, and a deep spiritual substance. These folk-art forms and artistic acts, precipitated by history, are vital to the preservation of cultural diversity and the direction of human development [1–3]. The establishment of a ceramic cultural repository can better protect and transmit ceramic culture . The creation of a ceramic culture resource bank can provide a platform for college and general ceramic industry teachers, students, ceramic practitioners, and enthusiasts to focus on accessing, studying, and researching ceramic-related literature resources, as well as providing reference, advisory, and service functions for the ceramic research industry, which can promote local economic and cultural development [5, 6]. The construction of a ceramic cultural resource library is a long-term project that necessitates the collaboration of many people. Problems will inevitably be encountered in all aspects during the construction and subsequent construction process.
China has a long history of ceramic culture and every inch of land and every corner of China has a strong ceramic cultural atmosphere, so, it is difficult to collect a wide range of ceramic cultural resources. The quality of the China ceramic culture resource base depends on the quality of the various types of ceramic culture resources . The construction of a distinctive resource library is not just about grouping documents or data collections, but also about providing readers with a complete system, so, that the distinctive resource library is a complete system of distinctive resources as possible, in terms of both time and geographical considerations. In the process of collating and processing ceramic cultural resources, libraries need to spend a lot of time and effort to ensure the immediacy, completeness, accuracy, and authority of the resources in the shape of a huge and uneven amount of resources information, so, it is difficult to collate and process all kinds of ceramic cultural resources.
With the advent of the era of big data, data mining technology will be more widely used and popular in most industries, and the information obtained through data mining technology can be widely used in various fields. The information obtained through data mining technology can be used in a wide range of fields, directly contributing to the information technology (IT) and intelligence of many industries.
We use data fusion and feature extraction approaches to do big data mining of ceramic cultural resource data. Semantic segmentation processing methods are used to identify, gather, schedule, manage, and aggregate heterogeneous ceramic cultural big data. It leads to the improvement of the information distributed storage and detection capability of ceramic cultural resource data. In the process of network construction, it is necessary to perform semantic segmentation of images, realize the quantitative analysis of ceramic cultural resources big data features, construct an optimized mining model of ceramic cultural resources, combine fuzzy correlation analysis methods for ceramic cultural resources big data mining and feature extraction, and analyze ceramic cultural resources big data information transmission . The ceramic cultural resources big data mining model is very important for the construction and design of ceramic cultural resources repository.
The basic goal of this paper is to create an efficient data mining and feature extraction model for ceramic culture large data using a fuzzy correlation analysis approach. A fine mining method for ceramic culture big data based on semantic segmentation is also proposed. The chunked region fusion and the semantic segmentation methods are used for the specific positioning of ceramic culture big data and fuzzy information feature extraction respectively. Similarly, the fuzzy semantic feature rule set is established for the optimal control of ceramic culture big data mining, the statistical information analysis method is used to establish the fuzzy feature distribution set for ceramic culture big data mining, and the semantic segmentation is used under the decision tree model for ceramic culture big data mining, to achieve data mining optimization. This study also reveals that the accuracy of data mining technique’s is high as compared to the earlier approaches, and the histogram distribution of the data mining method shows that the feature convergence is promising, and the data mining method’s anti-interference ability is strong.
The rest of the paper is organized as follows. Section 2 is about the related work which recaps some of the related work already existing in the literature on ceramic culture mining. The proposed methods for big data mining of ceramic data are discussed in Section 3 of the paper. Section 4 discusses the experimental result and analyses, in which the result of the proposed method is computed, examined, and compared with existing work. Finally, Section 5 is about the conclusion, which recaps the overall theme of the paper.
2. Related Work
Because of its unique status, Chinese ancient ceramics play an essential role in ancient China’s cultural history, and research on the digitization and retrieval of ancient ceramics has gotten a lot of attention as a consequence of major craft, cultural, and valuable goods. Zhang et al.  obtained the four texture features of roughness, contrast, directionality, and gradient by segmenting digitally archived porcelain images, using principal component analysis (PCA) and open-close operations combined with filtering and watershed algorithms to obtain the segmentation results of the main ceramic texture features. This method is an optimal exploration of segmentation specifically for porcelain images and establishes a certain basis for the digital management of museums, as well as the retrieval of texture features of porcelain.
Delanoy et al.  provided effective support for the computer-aided restoration of broken artifacts by dividing a subset of ancient ceramic artifact fragments, reducing the scale of exhaustion in fragment stitching, decomposing the difficult contradiction of forward stitching NP, and incorporating color-emotion features for efficient fragment stitching of ancient ceramic artifacts. The method is a typical machine intelligence-assisted heritage restoration and is an application of deep learning image classification in the restoration of ceramic artifacts. The method can assist restorers in achieving rapid artifact fragment classification and assemblage determination.
Mohr  used a digital image processing method in conjunction with pattern recognition techniques and provide an in-depth discussion of the automatic classification and identification of celadon ornaments using the special characteristics of the glaze of blue and white porcelain and the deep features such as color and texture of the digital image of decoration. Further, they used the K-nearest neighbor (k-NN) and support vector machine (SVM) classification methods to achieve the automatic classification of the decoration features of blue and white porcelain.
Song et al.  utilized the convolutional neural networks (CNN) and handwritten Chinese characters to segment and recognize the Chinese characters of blue-and-white porcelain inscriptions, achieving a preliminary recognition of the characteristics of ancient ceramic inscriptions. However, there are obvious limitations in the recognition method of segmenting the inscription mark image and then recognizing the Chinese characters, which cannot achieve effective recognition of the overall features of the inscription mark. Xu et al.  refined the structure data of Chinese archaic ceramics by measuring, recording, and computer coding and established a database of the structure of Chinese archaic ceramics, which facilitated the identification of some archaic ceramic features; however, the data volume was small and the coverage was incomplete.
Qin et al.  achieved digital quantitative analysis and structural feature extraction of bowl-shaped features by changing the status quo of relatively single and vague bowl-shaped descriptions, and applying the method as a new auxiliary analysis method, providing a new scientific basis for the identification and value recognition of ancient ceramics. This method mainly focuses on the measurement and collection of vessel types and does not generalize it to a single vessel type. Zhang et al.  proposed a machine inspection method by combining the Kirsch operator and Canny operator for surface defect detection and 3D reconstruction of ceramic bowls instead of manual inspection and designed a simple sorting system, which achieved a sorting accuracy of 95.3%. They proposed a 3D reconstruction matching method based on multiple image sequences, obtained spatial shape information through local point cloud reconstruction, added surface features points, and finally obtained the results of the overall reconstruction of the surface, which has certain significance for the restoration of some ancient ceramic relics.
Hadji et al.  explored the digital extraction method of structural features of ancient ceramic vessels, taking the skewed bowls produced at HuTian Kiln in Jingdezhen during the Five Dynasties, Song, and Yuan dynasties as an example. They improved the accuracy and effectiveness of the traditional experience-based visual sensory identification model of ancient ceramic artifacts based on image enhancement, edge extraction, and curve fitting by making full use of mathematical methods. It also combines the current methods of compositional data analysis and thermal light technology, which are gradually being applied and promoted in the field of scientific and technological identification of ancient ceramics, to further improve the comprehensive scientific and technological identification system of ancient ceramic relics.
Sadouk et al.  proposed a traversal of the edge contours of ancient ceramic vessels as the acquired ancient ceramic vessel features, and the multi-channel histogram under the Hue, Saturation, and Intensity (HIS) color space as the ornament features, and used a ML method to initially realize the research of lossless classification. This method has some shortcomings, as it does not fully take into account other features of ancient ceramics and does not have generalization characteristics. The vessel type and ornamentation features in this case are relatively simple, which has some reference value for this research. Xie et al.  scanned and then digitized the vessel types of restored YaoZhou kiln bowls excavated in different historical periods using the slicing method and used multivariate statistical analysis to clarify the indicators for classifying vessels types from the Tang, Five Dynasties, and Song dynasties.
In this paper, an association feature extraction method is adopted for fuzzy correlation analysis of ceramic culture big data, a cloud information processing platform is used for feature detection of heterogeneous big data, and an autonomous mining algorithm of ceramic culture big data is constructed. A fuzzy information clustering analysis method is utilized for numerical information group analysis of ceramic culture big data, and a joint analysis technique of point, line, and surface-like elements is adopted. We also establish an information attribute chain model, a fuzzy decision model for ceramic culture big data mining, and an adaptive optimization-seeking technique for optimal mining of ceramic culture big data. The attribute chain table of ceramic culture big data is established, and a spatial distribution structure model of ceramic culture big data is obtained.
3.1. Data Clustering Analysis
The adaptive weighted learning model of ceramic culture big data is built using the fuzzy information clustering model of ceramic culture big data, the feature matching technique, the fuzzy feature clustering of ceramic culture big data, and the semantic segmentation model. The fuzzy clustering distribution of ceramic culture big data is expressed bywhere is the association estimate of ceramic culture big data; is the measured value of ceramic culture big data collected at point ; is the distance between the two points and 0; S is the measured statistical feature amount of the measured points; and k is the interpolation weight of ceramic culture big data mining. According to the adaptive weighting learning results, the adaptive weighting of spatial features of ceramic culture big data is carried out, the fuzzy weighting learning formula of ceramic culture big data is constructed to improve its adaptive mining ability, and the spatial clustering model is defined aswhere x, y, and z denote the semantic similarity feature quantity of ceramic culture big data in three-dimensional space, σ denotes the semantic ontology set, b denotes the fuzziness coefficient of data mining, and r denotes the rough feature matching set.
3.2. Big Data Mining Optimization
In order to achieve feature extraction and information mining of big data, a neural network training algorithm, which is constructed by the Kohonen network [9, 10], is used. It is combined with an effective mathematical model of classification to establish a self-organizing feature mapping algorithm for unsupervised learning of big data mining. The topology of the Kohonen neural network for big data mining constructed in this paper is shown in Figure 1. The Kohonen neural network is divided into 3 layers, including the input layer, the implicit layer, and the output layer, where the upper layer is the output layer contact (set to N), and the integrated query interface generation module realizes the output of the mining feature information of big data, arranged into a feature space of nodes according to the distributed line column structure, with the input nodes at the bottom, realizing the input of the original data information features. The integrated query interface generation module enables the discovery of the Web database and the extraction of query excuse patterns.
Learning training for big data is performed according to the big data mining Kohonen neural network model shown in Figure 1. The steps of the training model for big data mining are described as follows:(1)Give the N number of classification and query interfaces (i.e., vector patterns) of the Web database and the k numbers of input nodes (i.e., each vector element) of the big data mining Kohonen neural network, and initialize the adaptive weighting vector from node I to output node j of the output layer of the Kohonen neural network. Set the adaptive weighting coefficients of the query result processing module to the number of random vectors, let us look at the sequence of training, and set the initialized pointer count t = 0 for the big data mining process.(2)Input the number of Web database size samples in the neural network topology for Kohonen big data mining, that is the training vector pattern.(3)Calculate the distance between the search query results and all big data clustering center join weight vectors, expressed as Euclidean distance in equation (1).(4)Find the dynamic inertia weight nodes of the neural network.(5)Adjust the differential evolutionary sequence of the big data clustering center vector and the output node for dynamic feature matching with the geometric neighborhood, where the adaptive weighting weights are shown in equation (2). Further, the present generation learning rate in the evolutionary process of the neuron for big data mining has the ability of similar form matching.(6)If the feature samples of big data mining are also input, then set and go to step 2.(7)Otherwise, the training is finished. Through the above processing, a neural network training model for big data mining is constructed, which is used as a basis for big data feature extraction.
Statistical detection of ceramic culture big data is achieved by mining a large number of semantic association features combined with the fuzzy attribute feature detection approach. With the use of statistical-analysis technique, a semantic-segmentation model of ceramic cultural big data is established , and the computational formula is given as follows:where is the global weighted value of the point of ceramic culture big data mining and denotes the clustering center of ceramic culture big data distribution nodes. The feature extraction model of ceramic culture big data is constructed, and data mining is carried out according to the feature extraction results. In the STARMA (1, 1) network model, the information ceramic culture big data visualization segmentation model is obtained, the output autocorrelation feature matching model of ceramic culture big data is constructed, and the statistical analysis of ceramic culture big data is combined with the fuzzy feature clustering analysis method . The semantic segmentation method is used for fuzzy information feature extraction, and at the feature point , the set of feature distribution at moment t is obtained and expressed as , where t denotes the number of ceramic culture big data and is the weighting coefficient of ceramic culture big data mining. Combined with the semantic feature analysis method, the fuzzy semantic feature rule set was established by Zhang R et al. , and the adaptive weighting coefficient was obtained as follows:where is the fuzzy constraint feature quantity for ceramic culture big data mining to find the best . The pheromone concentration for heterogeneous multi-core platform search is defined bywhere , N denotes the dimensionality of the big data mining nodes, denotes the information embedding dimensionality of the I, denotes the association information of the sampled data of node I and node j using big data, and t is the sampling time interval. The ceramic culture big data feature extraction technique is used for the mining of big data.
3.3. Semantic Dynamic Segmentation and Mining Output
A semantic dynamic feature investigation model of ceramic culture big data is established to extract the number of statistical features of ceramic culture big data for adaptive optimization seeking of ceramic culture big data mining  and is calculated aswhere and are the fuzzy rule feature quantities for ceramic culture big data mining. The numerical information investigation method is used to inaugurate the fuzzy feature distribution set for ceramic culture big data mining, which is expressed as follows:where m and n denote the number of sample embedding dimensions and the number of segmentation grids, is the statistical magnitude of the useful information of ceramic culture big data to be mined, and is the statistical average of the data, and n(t) is the interference term. In summary, the precise mining model of ceramic culture big data is constructed aswhere is the association dimension of semantic segmentation, denotes the number of output features to be mined, and denotes the semantic information component of data mining. Using the decision tree (DT) algorithm, semantic segmentation is used for adaptive optimization seeking in the process of ceramic culture big data mining. The statistical feature quantity is extracted, and the fuzzy C-mean clustering method is used for big data clustering according to the feature extraction results [21, 22], to realize the data mining of ceramic culture big data.
4. Experimental Results and Analysis
Edge detection is mostly used with grey-scale images, in which each pixel’s information is defined by a quantized grey level in a digital image. The improved edge detection algorithm adjusts the image grey values to distribute the whole interval, making the image contour direction, edge trend, and streaks very obvious and clear. Figure 2 shows a randomly selected ceramic data image as the experimental image.
As a result of the model being trained with the normalized value, the projected value of the ceramic product placement is likewise less than 1. Therefore, the predicted result needs to be inversely normalized; that is, the original normalization formula is inversely operated to obtain the real value, and then the four corners of the Bounding Box are re-calculated and connected. As shown in Figure 3, the Bounding Box encloses all the patterns inside the ceramic mug, and the remaining interference with the background information is basically outside the Bounding Box.
Image recognition with image cutting effect is the recognition of image features, which can be affected when there is too much background information. With the ceramic product positioning technique, we are able to know the exact position of the ceramic product in the image and cut the input image along the edges of the Bounding Box. This technique also enables us to remove any background information from the input image that is not relevant to the decorative pattern of the ceramic product.
As shown in Figure 4, in addition to the ceramic plate there is also extraneous information such as books and desktops in the image. Although the ceramic plate occupies the middle of the image, the proportion of the image it occupies is relatively small. Therefore, when the subject information only occupies a small part of the image, the accuracy of the image matching is greatly reduced.
The image shown in Figure 5 is a cut of the original image after the ceramic vessel type has been positioned. It is clear from the image that the ceramic plate takes up most of the area in the image and retains all the information, so there is particularly little interference.
In this paper, the ceramic vase image shown in Figure 6 is compared with the algorithm provided in the literature and the algorithm proposed in this section. Figure 6(a) shows the image foreground obtained from the ceramic vase image after edge segmentation, and the 2D grey-scale image obtained after trilateration; Figure 6(b) illustrates the 2D contour image after edge detection with feature points marked; Figure 6(c) demonstrates the 3D reconstruction effect obtained by the algorithm in the literature, and the resulting surface has protruding points and is not continuous, with poor smoothness. On the other hand, Figure 6(d) shows the 3D reconstruction result of the algorithm in this paper; the recovered surface is smoother in comparison, especially at the feature points where the recovery accuracy is higher. Intuitively, from the simulation results, the 3D model generated by the method provided in this paper is smoother in effect than the models generated in the literature and improves time efficiency while ensuring a high degree of feature information restoration. The algorithm in this section has some accuracy advantages in terms of reconstruction effect, reducing some of the image loss caused by reconstruction and reducing errors.
Python is used as a language for the simulation testing and analysis in order to validate the performance of the proposed method for performing ceramic culture big data mining. Different parameters setup during the experiment is given in Table 1 and described as follows: there are 120 IoT nodes sampled for ceramic culture big data, 12 data mining root nodes, 5 spatial dimensions of ceramic culture big data distribution, and 10 data clustering attribute categories. The initial frequency of sampling big data is set to f1 = 1.5 Hz and the final frequency is set to f2 = 2.3 Hz. According to the above simulation parameters, ceramic culture big data mining is carried out and the time domain distribution of data samples is obtained as shown in Figure 7.
The histogram distribution of the data mining is shown in Figure 8. The analysis of Figure 8 shows that the method of ceramic culture big data mining has great feature convergence and improves the anti-interference ability of data mining. The accuracy of data mining was examined, and the comparative results are displayed in Figure 9, indicating that this approach has a high level of accuracy. Table 2 summarizes the overall statistics of Figure 9.
From Figure 9, it is very clear that our proposed method performed really well and the accuracy of the proposed data mining approach increased with iteration; i.e., the accuracy was recorded as 91% for 100 iterations, 92% for 200 iterations, 94% for 300 iterations, 98% for 400 iterations, and 99% for 500 iterations.
Ceramic culture is based on an immaterial form of fine craftsmanship, a distinct way of thinking, and a profound spiritual substance. Ceramic culture has a long history in China, and it is the Chinese people’s spiritual home; therefore, it is very vital to protect it. It may be better protected and transmitted by establishing a ceramic cultural repository. Data mining image collection and processing is necessary to acquire ceramic-related data because the ceramic culture resources base has the biggest number of photographs, videos, and other materials. As a result, we used data fusion and feature extraction methods to do big data mining of ceramic cultural resource data, because it plays a very crucial role in the creation and design of a ceramic cultural resources repository. Further, for data information management, scheduling, identification, collecting, statistics, and aggregation of heterogeneous ceramic cultural big data, semantic segmentation processing methods are applied. A semantic segmentation-based fine mining approach for ceramic culture large data is also presented. Python, a general-purpose programming language, is used to carry out the experimental setup and perform simulation. The accuracy of the proposed data mining algorithm was examined, which indicates that this approach has a high level of accuracy as compared to existing work in the literature. Along with this, the information distributed storage and detection capability of ceramic cultural resource data is also improved.
The data used to support the findings of this study are available from the author upon request.
Conflicts of Interest
The author declares no conflicts of interest.
This paper is one of the phased achievements of 2020 Jiangxi University Humanities and Social Sciences project “the application of ceramic cultural resources in cultural and creative products from the perspective of authenticity, taking Jingdezhen Museum as an example” (project number: jc20227).
T. Groueix, M. Fisher, V. G. Kim, B. Russell, and MathieuAubry, AtlasNet: A Paper-Mache Approach to Learning 3D SurfaceGeneration, CVPR, London, UK, 2018.
A. Bernhardt, A. Pihuit, M.-P. Cani, and L. Barthe, “Matisse: painting 2D regions for modeling free-form shapes,” in Proceedings of the SBIM, pp. 57–64, Gejiu, China, 2008.View at: Google Scholar
C. Niu and M. Zhang, “Using image feature extraction to identification of ancient ceramics based on partial differential equation,” Advances in Mathematical Physics, vol. 2022, Article ID 3276776, 12 pages, 2022.View at: Google Scholar
J. Delanoy, A. Bousseau, M. Aubry, P. Isola, A. Alexei, and Efros, “What you sketch is what you get: 3D sketching using multi-ViewDeep volumetric prediction,” 2017, arX iv:1707.08390.View at: Google Scholar
R. Mohr, “Local grayvalue invariants for image retrieval,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19, no. 5, pp. 530–535, 1997.View at: Google Scholar
G. Song, J. Cheng, and K. Grattan, “Recognition of microseismic and BlastingSignals in mines based on convolutional neural network and StockwellTransform,” IEEE Access, p. 1, 2020.View at: Google Scholar
D. Xu, “Pointfusion: Deep sensor fusion for 3d boundingbox estimation,” in Proceedings of the IEEE Conference on Computer Vision andPattern Recognition, pp. 244–253, Huiyang, China, 2018.View at: Google Scholar
Y. Qin, J. Li, and P. Jiang, “Image stitching by feature positioning andseam elimination,” Multimedia Tools and Applications, vol. 2021, no. 3, pp. 1–13, 2021.View at: Google Scholar
I. Hadji and R. P. Wildes, “What do we understand about convolutional networks,” 2018, arX iv:1803.08834.View at: Google Scholar
L. Sadouk and T. Gadi, Convolutional Neural Networks for Human ActivityRecognition in Time and Frequency-Domain, 2019.
C. Xie and A. Kumar, “Finger vein identification using Convolutional NeuralNetwork and supervised discrete hashing,” Pattern Recognition Letters, vol. 119, pp. 148–156, 2019.View at: Google Scholar
R. Zhang, J. Tao, and H. Zhou, “Fuzzy optimal energy management for fuel cell andsupercapacitor systems using neural network based driving pattern recognition,” IEEE Transactions on Fuzzy Systems, vol. 27, no. 1, pp. 45–57, 2018.View at: Google Scholar
Buduma NN. Locascio, Fundamentals of Deep Learning: Designingnext-Generation Machine Intelligence Algorithms, O’Reilly Media Inc, Jindezhen, China, 2017.