Analysis and application of martial arts video image based on fuzzy clustering algorithm

Abstract

The intelligent scheduling algorithm for hierarchical data migration is a key issue in data management. Mass media content platforms and the discovery of content object usage patterns is the basic schedule of data migration. We add QPop, the dimensionality reduction result of media content usage logs, as content objects for discovering usage patterns. On this basis, a clustering algorithm QPop is proposed to increase the time segmentation, thereby improving the mining performance. We hired the standard C-means algorithm as the clustering core and used segmentation to conduct an experimental mining process to collect the ted QPop increments in practical applications. The results show that the improved algorithm has good robustness in cluster cohesion and other indicators, slightly better than the basic model.

Keywords

Fuzzy clustering algorithm membership martial arts video image segmentation intersection clustering object extraction

1 Introduction

With the rapid development of network and multimedia technology, a large number of martial arts and national fitness materials are stored in various fitness guidance systems in the form of videos and pictures. To better promote public fitness, easy to learn, and view, there are various needs for editing, segmentation, and integration of martial arts sports video. People pay more and more attention to martial arts and video processing is gradually deepening in the field of martial arts. The segmentation of athletes in martial arts video has become a research hotspot in digital image processing. In the past, many motion object segmentation algorithms are difficult to achieve good results in martial arts video, which makes the segmentation effect and real-time performance not very ideal. Because of the shortcomings of the existing motion object segmentation, speed is slow and the segmentation accuracy is easily affected by the irregular movement and illumination of the object, this paper also proposes a method of moving foreground object in martial arts video sequence based on frame difference intersection clustering Take the algorithm. Through the extensive simulation experiment in MATLAB 7.0 environment, the experimental results show that the algorithm is effective, the tube is feasible, the amount of calculation is less, and can get a more satisfactory effect. Video compression technology has gradually become a hot research topic. Video compression technology reduces the video capacity by video coding, which is convenient to release storage space and reduce the transfer and transmission time. The traditional video coding standards are MPEG2, H.263, and so on. MPEG refers to the moving picture experts group, representing the dynamic image expert group. Its computational redundancy is small, the overall performance is relatively stable [1], and the efficiency of H. 263 is higher. Besides, video coding technology needs to segment the object image to ensure the coding integrity of the object. However, traditional video coding standards can not search on video content, so it is difficult to segment images. MPEG4, which was launched in 2000, adds the semantic search function of video content, which can segment the image background and foreground into different semantic objects. The coding efficiency is improved, but it can not quickly eliminate the noise in the coding process. Time-domain segmentation and frequency domain segmentation are the first methods for motion video image segmentation. After many times of practice, these two methods can not accurately describe the pose of the target object, and the image segmentation is not clear.

Therefore, the research based on a fuzzy clustering algorithm to study the martial arts video with strong movement posture changes will help to better use and analyze martial arts video images.

2 Video-based training assistant system

In 1996, Semyon s. siobounov of Pennsylvania State University researched and developed the “computer graphics for improvement of springboard diving” system, which is used to help sports coaches and athletes to enhance their understanding of the whole body posture in the process of sports. “Aijie motion image measurement and analysis system” developed by Beijing Aijie human body information research institute uses multiple cameras to capture athletes’ technical movements, analyzes their three-dimensional motion images synchronously, synthesizes the three-dimensional space setting marks of joint points and human body center of gravity and calculates the athletes’ three-dimensional speed, human body posture angle, joint angle, etc. Although the data acquisition system will play a better role in the training of athletes, it will inevitably affect the effect of these sensors in the training system. At the same time, the system developed by them has high computational complexity and expensive equipment, which is not suitable for promotion. With the development and application of digital video technology, it is possible to obtain digital video images of athletes’ training and use video processing methods to obtain human motion parameters. In recent years, the research on video-based training assistant systems has emerged at home and abroad. “Video-based diving analysis system” of Tsinghua University, a key project of the Ministry of education. The idea of the project also focuses on the extraction of the moving objects in the diving video, and the extracted moving objects are used for video synthesis to eliminate the background influence. Besides, to make the whole image look more natural, some image enhancement techniques are used to extract the moving object boundary. In terms of several existing foreign software systems, different developers have different ideas about the emphasis of the system. Among them, dartfish’s video-assisted training and guidance software system and Simi company’s software all focus on video playback. The method is to fuse two video fragments of different actions into one video, and perform synchronous comparative playback (action comparison), and display the athletes’ movements at different times in the same background with panorama. The motion monitor system of ist company mainly aims at the demand of motion mechanics. It uses the fine human skeleton model and uses special data acquisition equipment to obtain the mechanical parameters of human motion.

3 Fuzzy clustering algorithm

The fuzzy clustering algorithm comes from the pattern clustering theory. It is an algorithm that uses mathematical rules to describe the segmentation interval. When the martial arts video image is segmented, the fuzzy clustering algorithm uses the iterative operation to segment the target object pixels and divides the image pixels into different membership intervals to make image segmentation decisions. When the target object has pixel samples, the pixel samples are formed into an unclustered set, which is represented as the representative pixel samples. Let denote the number of image segmentation types and the fuzzy clustering weighting factor. The image segmentation region is numerically equivalent to the weighted accumulation of fuzzy clustering of the sum of the squares of the distance between pixels and the clustering center, i.e., where: represents the set of membership functions, is the distance from the pixel point to the midpoint of the cluster; represents the type of membership of the unclustered pixel sample is the class. The set containing all kinds of membership is a fuzzy set of pixels, which represents the membership value of the sample [2]. From the membership value, we can see what kind of segmentation region the sample belongs to. At that time, it must not belong to the fuzzy set. At that time, it must belong to the membership value. The segmentation operation of the video image is extremely rigorous. Once the correct fuzzy set and the appropriate membership function are determined, the segmentation error can be eliminated.

4 Segmentation method of martial arts video image based on fuzzy clustering algorithm

The motion posture of the target object in the martial arts video is random, the changing trend of the image is fuzzy, and the segmentation region is not easy to determine [3]. The proposed martial arts video image segmentation method can segment the motion posture region of the target object and the useless pixel area into the foreground and background, and extract the movement situation of the target object. The fuzzy clustering algorithm extracts the image sequence under the playing state of martial arts video and starts the edge detection process at the same time. According to the image sequence, the motion prediction and compensation of the target object are carried out. The same background is established on the Ruo thousand images with a small distance. The time-domain difference image is extracted by using the attributes of different values of adjacent image frames, and the fuzzy property of the target object in the motion process is set After that, fuzzy clustering is carried out on the time-domain difference image. The edge contour of the target object is cut through the edge detection process to obtain the foreground region, and the background region is obtained by eliminating the foreground region.

4.1 Discussion on segmentation of fuzzy clustering algorithm

For the selection of fuzzy attributes in fuzzy clustering algorithm, we need to discuss the following two issues: one is which attribute of the target object can accurately describe the motion posture; the other is that the fuzzy attributes in the motion pose region and the useless pixel region cannot be the same, and there must be obvious differences. After the discussion, the proposed martial arts video image segmentation method finds that when the target object in the motion video image has undergone motion prediction and compensation, the background area can be represented as the image background difference with smaller spacing and the same background. At this time, if the background of the original martial arts video image has a normal distribution, then the segmented image should also have normal distribution And the average value of pixel distribution is equal to 0. Because the motion pose of the object in the motion pose region is random, the pixel distribution of the foreground of the segmented image cannot be determined directly, and the foreground region cannot be extracted, so it is necessary to measure the gray level of the image. The gray difference of pixels in different frames in the background is small, but the gray difference of pixels in different foreground images must be greater than that of background [4]. According to this property, the gray value can be taken as an important attribute when selecting fuzzy attributes. Besides, the first mock exam has a little gain on the segmentation effect. It can start from the normal distribution of background, and image segmentation process of moving pose region and useless pixel region as a process of describing non-normal distribution pixels in normal distribution pixels [5 –12].

The second-order matrix shows a smooth curve in the graph, which can optimize the image abnormal noise, and the image segmentation result is more reliable. There are four kinds of fuzzy attribute solution sets of the fourth-order matrix, which need dialectical selection. There are only two solution sets of the second-order matrix. There are many kinds of membership functions of the fuzzy clustering algorithm. Among them, s function can take into account the great difference of fuzzy attributes between foreground and background in the sports video image and can provide two kinds of membership degrees, i.e., maximum and minimum, which is easy to distinguish. Therefore, s function is used in foreground image segmentation. The background image noise is large and normal distribution, so it is necessary to select the membership function which can provide the time-domain variance. S function is in line with this point [13 –18].

4.2 Image extraction of sports video in playing state based on fuzzy clustering algorithm

The fuzzy clustering algorithm extracts the image sequence in the state of sports video playing and starts the edge detection process at the same time. According to the image sequence, the motion prediction and compensation of the target object are carried out. The same background is established on several images with small spacing, and the time-domain difference image is extracted by using the attributes of different values of adjacent image frames, After setting the fuzzy attributes of the moving object and writing the corresponding membership function, the time-domain difference image is fuzzy clustering. The edge contour of the target object is cut through the edge detection process to obtain the foreground region, and the background area is obtained by eliminating the foreground region.

4.3 Discussion on segmentation of fuzzy clustering algorithm

Because the motion pose of the object in the motion pose region is random, the pixel distribution of the foreground of the segmented image cannot be determined directly, and the foreground region cannot be extracted, so it is necessary to measure the gray level of the image. The gray difference of pixels in different frames in the background is almost the same, but the gray difference of pixels in different foreground frames is inevitably larger than that of the background. According to this property, the gray value can be taken as an important attribute when selecting fuzzy attributes. Also, the first mock exam has a little gain on the segmentation effect. It can start from the normal distribution of background, and image segmentation process of moving pose region and useless pixel region as a process of describing nonnormal distribution pixels in normal distribution pixels [27 –30].

The fourth-order matrix has a large amount of calculation and is insensitive to the normal distribution characteristics. Therefore, the fourth-order matrix is simplified to the second-order matrix to extract the fuzzy attributes of the time-domain difference image [27 –34].

5 A fourth-order matrix is used to extract non-normal distribution pixels, and a moving hollow rectangle is used to measure the time-domain matrix

5.1 High spatial accuracy

Generally speaking, the smaller the SA, the higher the spatial accuracy. The fuzzy clustering algorithm in this paper has the smallest SA value among the four segmentation methods, which can effectively fill the space gap of martial arts video, prevent the vacancy in the image segmentation area, and accurately segment the image with complex environment and color.

5.2 Noise iteration performance

To get a high-definition image, the segmentation method needs noise iteration to get a high-definition image. The basic requirement of noise iteration performance of the segmentation method is fast adaptive convergence, which is the noise average curve of two continuously changing moving images in martial arts video, and the relationship curve between the iteration times of different segmentation methods and the average noise value. The average maximum noise values of two continuous moving images are 0.40 dB and 0.23 DB respectively, and the overall rising trend of noise is relatively gentle. After the processing of the segmentation method, the image noise is reduced. In this paper, the martial arts video image segmentation method based on fuzzy clustering algorithm can carry out fast adaptive convergence under the same iteration times, and the image noise can be reduced to 50% of the original after 6 iterations, which has good noise iteration performance.

5.3 Spatial distortion rate

The spatial distortion rate of two continuously changing moving images in different segmentation methods is analyzed. The spatial distortion rate is related to the number of frames and the segmentation effect, so the spatial distortion rate is lower than that with fewer frames. Besides, the algorithm can reduce the noise of the moving object in the image segmentation. However, the other three segmentation methods have a high spatial distortion rate and unstable trend, which can not effectively deal with the image noise points, and the background construction is not successful, so the segmentation image is fuzzy.

Fuzzy clustering sets the number of clusters K when it is initialized. In the algorithm proposed in this paper, combined with the method of pedigree clustering (i.e., selecting different cluster numbers to cluster in turn), and taking the size of information entropy as the criterion to judge the number of clusters K, this method calculates the information entropy of each clustering state and selects the clustering state with the minimum information entropy as the clustering result.

The method of pedigree can solve the problem caused by users’ pre-set cluster number K. however when fuzzy clustering initializes the cluster center, the randomness is too strong, which may cause the number of clusters to deviate from the actual number (although this possibility is very small). Therefore, this algorithm combines the diverse setting of Fuzzy degree m and accuracy E, The majority voting method is used to remove the random deviation, and the result is very close to the actual clustering partition.

6 An algorithm of athlete foreground object extraction based on Clustering

6.1 Clustering algorithm based on cumulative frame difference intersection

The basic idea of the algorithm is to cluster the two kinds of accumulated frame differences based on the operation of cumulative frame difference, to make the k-domain converge to the foreground edge accurately. Then, the region is binarized to get the mask image of frame difference, that is, while keeping the time of confirmation, the mask image of frame difference is obtained, It greatly improves the segmentation effect, and is also suitable for the martial arts video sequence with fast speed off-moving object. The main steps of the algorithm are as follows:

6.2 After median filtering, the adjacent frame difference and inter-frame difference are performed for the images in the sequence

If the current nth frame image can be expressed as f (k), then the K + 1; and K + 2 images are represented as f (K + 1) and f (K + 2), then the adjacent frame difference and the inter-frame difference can be expressed as follows: $| f_{k + 1} (x, y) - f_{k} (x, y) |, | f_{k + 1} (x, y) - f_{k} (x, y) |$ (1)

Two kinds of cumulative results of the first N frames are as follows: $D_{1} = \sum_{k = 1}^{n} | f_{k + 1} (x, y) - f_{k} (x, y) | = E \cup N_{1}$ (2) $D_{1} = \sum_{k = 1}^{n} | f_{k + 2} (x, y) - f_{k} (x, y) | = E \cup N_{2}$ (3)

Among them, E and F are the various regions of cumulative results, and N₁ and N₂ are noise

6.3 The intersection of the two kinds of cumulative frame difference can effectively concentrate the pixels belonging to the change region in the two kinds of cumulative results around the foreground contour, and obtain the ideal moving foreground contour

$\begin{matrix} D_{1} \cap D_{2} & = (A \cup N_{1}) \cap (B \cup N_{2}) \\ = EF \cup ({EN}_{2} \cup {FN}_{1} \cup N_{1} N_{2}) \end{matrix}$ (4)

Since EN₂ and FN₁ are the clusterings of motion region and noise, formula (5) can be rewritten as follows: $D_{1} \cap D_{2} = AB + N$ (5)

Among them, EF represents the contour of moving foreground and N is noise.

6.4 The clustering method is used to remove the old background image points. The specific steps of clustering are described as follows

Two points were randomly selected as the initial clustering centers, and two rectangle windows with a size of 5x5 were set with the two points as the centers, and the number of cycles was set as 1p;

For each image point in the shape knowing window, the distance DL from it to the cluster center is calculated according to formula (6); $D_{l} = \frac{1}{N_{k}^{2}} \sum_{(x_{h}, y_{h}) \in h (x, y)} \sqrt{(x_{h}, x)^{2} + (y_{h}, y)^{2}}$ (6) Where, (x_h, y_h) is the coordinates of pixels whose gray value is not 0 in the rectangular window; NH is the number of pixel points in the rectangular window whose gray value is not 0.

Step 3. Compare the distance dl with the threshold D_t. If D_l < D_t, the pixel will be merged into the edge moving image of the moving foreground object; otherwise, the pixel will be merged into the background edge noise pixel;

Step 4. Move the rectangular window along the horizontal and vertical directions to increase the number of purple dots. Execute step 2, until all pixels are merged into the class, that is, the number of cycles reaches the specified number, and the clustering ends.

7 Global motion estimation

The global motion refers to a large proportion of pixel motion in the video sequence, which is mainly caused by camera motion. Usually, the video image is composed of foreground and background. If the camera is moving during the shooting process, and the foreground object also has its motion, then the background and foreground have their motion in the video sequence: the background motion is caused by the camera motion, which is called global motion; The motion of foreground object is the motion of foreground object relative to the camera, which is called local motion. The purpose of global motion estimation is to find out the rules of camera motion which cause global motion from the video sequence. Global motion estimation is widely used, such as sprite coding, global motion compensation coding, motion-based object segmentation, and so on. The global motion is mostly caused by camera motion. Researchers study the global motion law by modeling the camera motion. To model the motion of the camera, it is necessary to make some assumptions about the scene and the objects in it. In practical applications, data sources often contain massive data. When processing massive data, a sequential fuzzy clustering algorithm needs a lot of IV o overhead and enough memory space, which leads to the bottleneck of scalability and response time, Therefore, the establishment of a parallel fuzzy clustering algorithm is an urgent problem. The basic idea of a parallel fuzzy clustering algorithm is: massive data are stored in a parallel computer in blocks. When the algorithm is executed, the parallel computer only needs to calculate the corresponding block data and exchange the calculation results through the communication between parallel computers, so the computing time of the fuzzy clustering algorithm can be shortened. The surface of an object in three-dimensional space is usually regarded as a combination of many spatial planes, each of which satisfies the following constraints: $a X + bY + cZ = 1$ (7)

Let (X, Y, Z) and (X^″, Y^″, Z^″) be the three-dimensional coordinates of a point on an object in two adjacent frames. mutually

The coordinates of the points on the corresponding image plane are (Y, X) and (X^″, Y^″). If the movement of objects in a three-dimensional scene

If the motion is displacement, rotation, and linear change, then the relationship between (X, Y, Z) and (X^″, Y^″, Z^″) is as follows: $[\begin{matrix} X \\ Y \\ Z \end{matrix}] = [\begin{matrix} s_{11} & s_{12} & s_{13} \\ s_{21} & s_{22} & s_{23} \\ s_{31} & s_{32} & s_{33} \end{matrix}] • [\begin{matrix} X \\ Y \\ Z \end{matrix}] + [\begin{matrix} t_{1} \\ t_{2} \\ t_{3} \end{matrix}]$ (8)

If the motion of an object with the above surface characteristics satisfies this formula, we can obtain a perspective model with eight parameters or an affine model with six parameters by using the perspective projection method or orthogonal projection method. By further simplifying the affine motion model, the rigid model and displacement model can be obtained.

7.1 Pretreatment

To improve the accuracy and efficiency of video moving object segmentation, reduce the impact of noise, and enhance the effect of frame difference calculation, it is necessary to denoise the initial video sequence. This algorithm uses the most commonly used 3-3 square module to preprocess the median filter. It can quickly bite the line, protect the image edge, effectively remove noise. At the same time, it can also save the loss of image details, optimize the image quality, and lay a good foundation for the next step segmentation.

7.2 Binarization

To eliminate the non-zero difference of the frame difference image caused by the internal noise in the image, this paper uses the least half – sample method to calculate the threshold required by binarization.

7.3 Morphological treatment

7.4 Experimental result

n order to verify the effectiveness of the algorithm, this paper selects a long jump video image sequence to simulate the algorithm in MATLAB 7.0 environment. The test image sequence in Fig. 1 is a video from the long jump of martial arts athletes. These images are shot at the speed of 25 frames / s, and the size of the images is 640 x480. The results show that the algorithm can effectively extract the athletes in the foreground. The effect of accumulated cross-frame processing is shown in Fig. 2.

Fig. 1

Two frames in the original sequence image.

Fig. 2

The results are captured after processing such as cumulative frame difference intersection clustering.

Data set is a data set specially used to evaluate the performance of the clustering algorithm. Its data volume is more than 60000. To facilitate visualization, only a subset of 200 data is taken to verify the performance of the algorithm. Fig. 1 shows the display results of the FCM algorithm on the data subset. It can be seen from the figure that the FCM algorithm can be divided into three categories (three types are represented by symbols in Fig. 1), Moreover, the distribution of category center and error number obtained by running the FCM algorithm for many times is quite different. This is because the initialization center value is different, resulting in the final convergence position always slightly different. After using the improved algorithm, when the number of classes is 2, the information entropy value is the minimum. After running the new algo-rithm for many times, the result is relatively stable, and the information entropy is the minimum when it is divided into two categories, The new algorithm objectively reflects the actual situation.

Most of the existing clustering algorithms, both hard clustering algorithm and fuzzy clustering algorithm, adopt the method given in advance to solve the problem of determining the number of clusters. This paper combines information theory to connect the attribute of data point pair clustering with information entropy, and through the experimental table, the new algorithm can better solve this problem.

When the number of processors is 4, the relationship between the size of the data block and the response time is given. The results show that the parallel algorithm has a sublinear performance, and with the increase of the size of the data set, the parallel algorithm becomes more effective, The proportion of communication cost in the total time spent will be reduced.

7.5 Information entropy

In statistical physics, thermal entropy is a measure of the disorder of a physical system. The more ordered and certain a system is, the smaller the thermal entropy of the system is. In information theory, information entropy is a measure of the amount of information contained in a message sent by a source. The more certain a message is, the smaller the information entropy is.

Entropy is used to describe the disorder degree of atom distribution. The distribution of data points is similar to the distribution of atoms. The more reasonable the division of clustering is, the more certain the attribution of data points on a certain cluster is, the smaller the information entropy of the cluster is. In clustering analysis, the subjective division of data points to a cluster depends on the algorithm selected by users, When users adopt different algorithms, the attribution of data points is different; objectively speaking, the attribution of data points to a certain cluster is determined. Therefore, if the attribution of data points can be determined as much as possible subjectively, that is to say, to obtain the clustering result with the minimum information entropy value, the purpose of clustering is achieved.

According to the expression of information entropy, 3, information entropy is required, as long as the probability of the occurrence of each signal can be obtained. In clustering, the probability can be understood as the probability that any sample belongs to a certain cluster. Fuzzy clustering uses the membership function U_ij to express the degree to which any sample I belongs to cluster J, which is different from the existing algorithms, present, FCM, PCM, FPCM are the typical fuzzy clustering algorithms. In this paper, the FCM algorithm is introduced as an example.

8 Conclusion

Compared with MPEG4, frequency domain segmentation, and time-domain segmentation, this method is in a leading position in the experimental evaluation of spatial accuracy, noise iteration performance, and spatial distortion rate. The application of the video image analysis method of sports will help to grasp the laws of sports technology and the characteristics of movement skills of fitness crowd through video image analysis and serve for the improvement of physical education, national fitness, and competitive sports. In this paper, an algorithm of capturing the motion foreground information of martial arts video sequence based on frame difference intersection clustering is proposed. In the simulation experiment, the result is satisfactory, Compared with the existing methods, the algorithm has a certain advantage in speed and can meet the real-time requirements. The experimental results show that this algorithm is a better algorithm for extracting foreground athletes from martial arts video sequences.

References

Lixin

, Xuzhi

, Xianrui

, et al., Image segmentation based on fuzzy mean clustering algorithm optimized by adaptive Drosophila algorithm [J], Control Engineering 23(4) (2016), 494–499.

Shaohua

and di LAN

L.J.

, Multidimensional fuzzy clustering image segmentation algorithm based on the kernel and local information [J], Computer Applications 35(11) (2015), 3227–3231.

Jie

and Honghui

, Research and Simulation of an improved fuzzy clustering image segmentation algorithm [J], Computer Simulation 32(4) (2015), 380–383.

An image segmentation algorithm based on weighted image segmentation algorithm is proposed in this paper

Jinzhi

, Jinmin

and Yulin

, Improved WFCM image segmentation algorithm based on an optimized initial clustering center [J], Software 36(4) (2015), 136–142.

Fang

and Yan

, Medical ultrasound image segmentation algorithm based on hyperpixel and fuzzy clustering [J], Semiconductor Optoelectronics 37(1) (2016), 146–150.

Bartoli

, Dalal

, Bore

and Horud

, from Video Sequences to Motion Panoramas [A]-, Proceedinsg of the workshop on Motion and Video Computing(Motion) 2002, 201–207.

Yanling

, Zhichao

, Zhenming

and Qingchao

, The automatic extraction method of moving objects in the surround image sequence Xiao, Computer Engineering and Science 28(2) (2006), 63–65.

Chuanjin

and lulu

, Node localization algorithm for wireless sensor networks based on fuzzy c-means clustering [J], Electronic Design Engineering 24(8) (2016), 58–60.

10.

Zhidan

, Hongjie

Y.Z.

, et al., Adaptive image inpainting algorithm based on the sparsity of block structure, Acta Sinica Sinica 41(3) (2013), 549–554.

11.

Jun

and Liji

, 3D reconstruction of building facade based on image matching and point cloud fusion [J], Acta Computer Sinica 35(10) (2012), 2072–2079.

12.

Kaiyu

and Deying

, infrared dim small target detection method undersea and air background [D], Journal of Harbin Engineering University 30(9) (2009), 996–1000.

13.

Guangbo

, Yong

and Qian

, Image compression is based on an improved vector quantization algorithm [J], Science and Technology and Engineering 10(14) (2010), 3517–3519.

14.

Stolen

, Preuss

, Stoean

, et al., Multimodal optimization employing a topological species conservation algorithm [J], IEEE Transactions on Evolutionary Computation 14(6) (2010), 842–864.

15.

Liang

J.J.

, Qu

B.Y.

, Mao

X.B.

, et al., Differential evolution based on fitness Euclidean-distance ratio for multimodal optimization [], Neurocomputing 137(8) (2014), 252–260.

16.

, Bo

, Rong

, et al., Image segmentation based on the correlation between transmittance space and color texture [J], Electronic Measurement Technology 38(1) (2015), 41–46.

17.

Yuzhuo

, Qi

and Yuwen

, The international standard for object-based multimedia data compression coding MPEG-4 and its verification model, Beijing, Science Press (2000), 10.

18.

Baiyang

, Chun

and Ying

, The development of video segmentation technology, Computer Research and Development 38(1) (2001), 36–42.

19.

Sonkaetal (translated by AI Haizhou and Wu Bo)

, image processing, analysis andmachine vision, Second Edition, Beijing, people’s Posts and Telecommunications Press, (2002). (Chinese version)

20.

Jie

L.L.

, Dejun

and Weng

S.M.

, Nan a motion-oriented video object segmentation algorithm, Acta Sinica 23(12) (2000), 1326–1331.

21.

Yaxiang

and Wenyu

, optimization theory and method, Beijing, Science Press, (1997).

22.

Yuwen

, Shiqiang

and Yuzhuo

, Feature point selection and robustness analysis in global motion estimation, Acta Computer Sinica 24(3) (2001), 236–241.

23.

Yuwen

, Li

, Yuzhuo

and Shiqiang

, Fast and robust global motion estimation algorithm, Acta Software 12(8) (2001), 1220–1228.

24.

Yuwen

, Yuzhuo

and Shiqiang

, A fast global motion compensation coding method, Acta Electronica Sinica 29(2) (2001), 175–177.

25.

Hong

, Image segmentation algorithm and application based on clustering and fuzziness [D]

26.

Aihua

, Research on image segmentation technology based on fuzzy clustering analysis [D], Huazhong University of science and technology

27.

, The fuzzy clustering algorithm is based on intraclass and inter-class distance and application of image segmentation [D]. (2016).

28.

Xianghui

, Research on image segmentation algorithm based on MRF and fuzzy clustering [D]. (2019).

29.

Min

and Jian

, The fuzzy clustering algorithm is based on partition [J], Acta Software Sinica 15(006) (2004), 858–868.

30.

Wenping

, Yuanyuan

, Hao

, et al., Image segmentation based on rough set and differential immune fuzzy clustering algorithm [J], Acta Software 2014(11), 002675–2689.

31.

Jun

, Applied research on the fuzzy clustering algorithm [D], Zhejiang University

32.

Yunlong

and Lingjuan

, Implementation and application of a fuzzy clustering algorithm based on spark [J], Computer Technology and Development 029(001) (2019), 130–134.

33.

Yuanzhen

, Jian

and Chenyang

, An improved fuzzy clustering algorithm [J], Journal of Huazhong University of science and Technology (NATURAL SCIENCE EDITION) 33(002) (2005), 92–94.

34.

Yanhong

, Research on image segmentation algorithm based on fuzzy clustering analysis [D]. (2007).