KR100598134B1 - Method and system for vector data compression using k-means clustering - Google Patents

Abstract

The present invention provides a vector data compression method and system using K-average clustering to efficiently reduce the size of storage space occupied by spatial data through K-average clustering and dictionary-based compression.

The present invention divides each geometric object into a sequence of starting point coordinates and differential vectors, and then separates the differential vectors into lengths and angles, and performs K-average clustering for the separated lengths and angles, respectively, to enter a cluster center. It provides a vector data compression method suitable for a mobile environment through the process of making a dictionary having a function and converting it into a pointer indicating an entry of the dictionary by approximating each length and angle.

Mobile, differential vector, K-means clustering, dictionary-based compression, vector data

Description

TECHNICAL AND SYSTEM FOR VECTOR DATA COMPRESSION USING K-MEANS CLUSTERING

1 illustrates an example of a line string for explaining dictionary-based compression.

FIG. 2 is a preliminary diagram of FIG. 1.

3 is a flow chart illustrating a general clustering process.

4 is a general flowchart of a vector data compression method using K mean clustering according to the present invention;

5 is an explanatory diagram for the case where the adjacent starting point is the same cluster in the present invention.

6 is a hardware block diagram for implementing the present invention.

7 shows experimental data in the present invention.

FIG. 8 is a diagram illustrating differential vectors for the experimental data of FIG. 7. FIG.

9 is an enlarged view of the differential vector excluding the starting point.

10 shows a vector dictionary which can be combined in an experiment of the present invention.

11 illustrates a vector dictionary according to a conventional compression method.

12 is a table for analyzing the results for the present invention.

13 to 15 is a view showing the results of the experiment while increasing the pre-size of the length and angle in the present invention, respectively.

16 is a study result analysis table for the conventional compression method.

17 is a graph showing the relationship between compression ratio and position error.

<Description of the symbols for the main parts of the drawings>

100 input unit 200 differential vector extracting unit

300: compression unit 400: control unit

500: storage unit 510: starting point coordinate DB

520: length dictionary DB 530: angle dictionary DB

540: length dictionary pointer DB 550: angle dictionary pointer DB

The present invention relates to a method and system for compressing data in a vector form used in a mobile environment, and in particular, a K-average for efficiently reducing the amount of storage space occupied by spatial data through K-average clustering and dictionary-based compression techniques. A vector data compression method and system using clustering.

Recently, the use of mobile devices such as mobile phones, PDAs, car navigation terminals, etc. is increasing rapidly. In these mobile devices, the use of spatial data is indispensable for route searching and map services.

However, these mobile devices still have limited computational performance and limited storage space compared to desktop environments. Therefore, even vector data, which takes up relatively little storage space compared to the raster data, is still a big burden.

The vector data used in the mobile environment includes data for distance measurement or route search and map data for background. In the case of map data, even if a slight position error is included, it can be accepted if the error degree is indistinguishable from the eye.

To date, the compression technique for vector data in the field of data compression has not been as diverse as that of raster data. A typical dictionary-based study on vector data compression was published in 2000, "Design Algorithms for Vector Map Compression (D. Salomon. 'Data Compression: the Complete Reference.'Springer-Verlag, 2nd edition, 2000.)". In this study, we propose a method to reduce the amount of data by dividing the vector data into individual differential vectors and approximating them in advance using pre-fabricated methods using FHM (Fibonacci, Huffman, and Markov). In this study, the process of making a dictionary can be omitted by using a pre-made dictionary, but since the dictionary cannot reflect the characteristics of the data, it has a lot of positional errors in restoration.

In another study, "Vector Map Compression: A Clustering Approach (Shashi Shekhar, Yan Huang, Judy Djugash, Changqing Zhou, 'Vector Map Compression: A Clustering Approach', Proceeding of the tenth ACM international symposium on Advances in geographic information systema, 2002, 74-80). The results of this study show better performance in terms of positional accuracy during restoration than the FHM method, but still have a disadvantage in that it is still not satisfactory when applied to actual data.

The present invention has been made in view of the above, and an object of the present invention is to create a dictionary by dividing a differential vector representing a relative position of each geometric object by a length and an angle and clustering each of them by applying a K-average clustering technique. The present invention provides a method and system for compressing vector data using K-average clustering to obtain an improved compression ratio while having a difference in position accuracy loss that is difficult to distinguish through minimal computation.

In accordance with an aspect of the present invention, a vector data compression method using K mean clustering comprises: separating each geometric object into a sequence of starting point coordinates and differential vectors; Separating the differential vector into length and angle; Performing a K-average clustering on the separated lengths and angles to produce a dictionary having an average value of the cluster as an entry; And approximating each length and angle to convert the pointer to a pointer indicating a dictionary entry.

In addition, the vector data compression system using the K mean clustering of the present invention, the differential vector extraction unit for separating each geometric object into a sequence of the starting point coordinates and the differential vector from the data for each geometric object input through the input unit; The differential vectors extracted by the differential vector extracting unit are separated into lengths and angles, respectively, and a K mean clustering is applied to produce a dictionary for each of lengths and angles. A compression unit that converts and compresses a pointer to a value; A storage unit having respective databases storing starting point coordinates from the differential vector extracting unit, a length dictionary, an angle dictionary, a length dictionary pointer, an angle dictionary pointer, etc., from the compression unit; And a control unit for controlling the respective units.

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings. However, the following examples are merely to illustrate the present invention is not limited to the contents of the present invention.

First, a look at the dictionary-based compression technique and the K-Means Clustering technique applied to the present invention.

The dictionary-based approach (Keith C. Clarke, 1990) to be used in the present invention is not an approach developed solely for spatial data. As a technique applied to various other fields, the present invention designs a vector data compression method using a dictionary-based approach.

The dictionary-based approach is to create a dictionary with entries of the data to be expressed and then represent the data as a list of pointers to entries in the dictionary, rather than storing values in the actual data.

Applying this to vector data is as follows. When a line string (LineString) exists in the two-dimensional space as shown in Figure 1, there are several ways to store it.

The method proposed by the standard of OGC (OGIS, 1999) saves the actual coordinates using an array of x and y having 8 byte size together with type information and the total number of points. 120, 100 180, 200 100, 260 140, 160 160, 220 240). The method of storing this in a dictionary-based method is as follows. First, a dictionary is produced as shown in the table of FIG. 2 having individual point coordinates of LineString as an entry.

In the table of FIG. 2, the coordinates of all points of the LineString are eliminated. LineStrings are not represented by actual coordinates, but rather by arrays of pointers to preconfigured entries. Therefore, the structure is stored as '(1, 2, 3, 4, 5, 6)' as follows. In fact, when you want to use this stored spatial object, it searches for the entry of the dictionary pointed to by the array that replaces the object coordinate and expresses it as coordinates.

This data structure is called a dictionary-based approach. As shown in the example above, even if you create a data structure using a dictionary-based approach to a single object, the size of the entire data is not reduced. Since the size of the pointer pointing to the entry is added to the entry constituting the dictionary, the total amount of data is increased by the size of the pointer rather than simply recording coordinates. The present invention proposes a method of reducing the amount of data by making a dictionary by making entries having a lot of similar forms of overlap for a plurality of spatial objects.

Next, we look at the K-mean clustering.

Cluster analysis is one of the data mining techniques that extracts and analyzes meaningful information from large databases. Clustering here is a set of similar data (Jiawei Jan, Micheline Kanber, 2000). In other words, clustering is such that one cluster contains similar data and the other cluster is a set of less similar data. Therefore, cluster analysis is an analysis method that classifies heterogeneous objects mixed into several homogeneous clusters by similarity.

Clustering process is a four-step process as shown in FIG.

In the first stage of variable measurement, variables are determined that measure the characteristics of each individual that can be used to cluster the individuals. That is, m variables are measured for n individuals. m is the number of variables in the object to be used for actual clustering. The second step is to measure the similarity, which is to calculate the distance or dissimilarity between the entire entities using the m measured variables. The method of measuring the distance between the objects may be calculated differently according to the clustering method. In this way, the distance between all the objects is calculated to form a distance matrix that represents dissimilarity. Dissimilarity, the smaller the value, the closer the two individuals are. Closeness between individuals means that the m variables measured for clustering in individuals have similar characteristics. In the third step, the dissimilarity matrix computed by the selected clustering method is used to group objects with close distance into one cluster. The final step of the analysis is the process of determining the nature and interrelationship of each community. This process is an analysis process to determine what the individual clusters mean and the relationship between the clusters through the result of the clustering process.

Cluster analysis can be divided into hierarchical clustering technique and non-hierarchical clustering technique.

Non-hierarchical clustering is a method of dividing a given n objects into k clusters using a predetermined partitioning method. Non-hierarchical clustering requires several conditions. The first is that k must always be less than n. The number of final clusters must be less than the number of individuals and each cluster must contain at least one individual. Finally, each individual must belong to only one cluster. A representative of the non-hierarchical clustering techniques is the K-average clustering technique.

Hierarchical clustering is a clustering technique that hierarchically decomposes a given set of objects. There are two hierarchical clustering techniques, a bottom-up method and a top-down method. Bottom-up method is to assign each individual to a different cluster, measure similarity, and merge the clusters with high similarity into one, and repeat until the final cluster is merged into one or satisfies a given condition. The top-down method is a method of clustering by repeating the process of dividing all the objects into one cluster and splitting the cluster until one cluster consists of only one object or satisfies a given condition. Representative hierarchical clustering techniques include single joining, full joining, mean joining, and Ward. Hierarchical methods have the disadvantage that they cannot be recovered if an initial improper merge or split occurs.

The K-average clustering technique to be applied to the present invention is a non-hierarchical clustering method developed by MacQueen (MacQueen, 1967). In other words, it divides n entities into k clusters and divides them into one cluster.

The first step is to select K cluster centers as input parameters and select K cluster centers. Initial cluster centers are chosen arbitrarily. The next step is to assign each individual to the selected cluster. In this step, similarity is calculated and assigned to the cluster with the highest similarity. There are several ways to calculate the similarity. The K-average clustering method calculates the center value of a new cluster using the average of the objects assigned to each cluster. For clustering of two or more variables, the center of cluster is recalculated using the centroid instead of the mean. This process is repeated until either the given conditions are met or there is no shift in the center of the cluster.

There are several ways to assign the second stage of the clustering process to the most similar cluster. Minkowski distance, Euclidean distance, standardized distance, and Mahalanobis distance can be used as tools to measure similarity between objects. Euclidean distance is the most commonly used measure of similarity in applying the K-average clustering technique. In the present invention, the Euclidean distance is used as a tool for measuring similarity because the computation must be performed using geometric coordinates in two-dimensional space. The equation for calculating Euclidean distance is extended to be applicable in space in n dimension.

Using this distance calculation, the similarity between the individual and the cluster is measured, and the cluster is recalculated by assigning the object to the cluster having the highest similarity with the individual. This process is repeated until reassignment does not occur or a given condition is met to obtain the final result.

The K-average clustering technique has the advantage of finding meaningful information without prior knowledge of the internal structure of the entire data (Anil K. Jain, M. Narasimha Murty, 1999). In addition, if the distance relation between the observed value and the cluster center is defined according to the data type, it can be applied to most types of data. Even if an object belongs to the initial wrong cluster, iteration is reassigned to a valid cluster. The K-average clustering technique is easy to apply because it does not require any prior information other than the initial value K. However, if the initial value K is not properly selected, satisfactory clustering results cannot be obtained. In the process of defining dissimilarity distances, it is difficult to define them as one distance if the measurement scale of different data types (Jay lee, David w. S. Wong, 2000) is different. There is also difficulty in interpreting the results because there is no prior clustering objective.

The present invention uses the lossy compression method using the K-average clustering technique for the map data used as the background in the mobile environment, and the loss rate of the data so that it can be actually used rather than the compression rate for the entire data through the cast-based approach. Compression in the direction of minimizing and increasing the position accuracy in relation to the compression accuracy and the positional accuracy due to data loss, the present invention based on this concept will be described with reference to the operation flowchart of FIG.

First, in order to apply the K-means clustering technique to compress vector data, a process of separating geometric objects into starting point coordinates and a sequence of differential vectors is required.

The differential vector is a vector representing a relative position by using a difference between the coordinate of the previous point or the coordinate of the point where the object starts to express the position of the current point. In the present invention, the difference cell vector is extracted using the difference between the corresponding coordinates of the individual coordinates and the previous coordinates (S100).

Then, the extracted differential vectors are separated into lengths and angles, and a dictionary is produced by applying K average clustering to each of them (S200 and S300).

At this time, clustering is performed on the length and angle of the differential vector using the K-average clustering technique, and the center is calculated by using the allocated objects in the cluster. In the present invention, since the two-dimensional differential vector is separated into one-dimensional length and angle, the center is calculated using each average value. A dictionary is created with entries of representative values in the cluster calculated in this way. The object is converted into a pointer indicating a representative value of a cluster to which the object belongs by using the prepared dictionary using a dictionary-based compression technique (S400).

Compressed data through this series of processes is the set of the starting point of the object representing the absolute position of the spatial object, the dictionary of the length and angle of the differential vector representing the relative position from the starting point, and the array of two pointers pointing to it. And so on.

In the above case, the start point of the spatial object is stored as it is without any processing. If the differential vector represents the relative position of individual point data, the starting point of the object is a value with an absolute position in coordinates. The spatial object can be reconstructed only when there is a reference for absolute position and a differential vector for each relative coordinate.

When the clustering is applied to the starting point as well, if the number of clusters is smaller than the number of objects, that is, the starting point, two or more polygons as shown in FIG. 5 are combined into one.

If the number of clusters is determined to be less than or equal to the total number of objects when clustering is applied together with differential vectors representing other relative positions even coordinates that can represent the absolute position of the object, the position error for the entire data may be reduced. Due to the same phenomenon, the data is severely distorted and becomes unusable data. If you apply clustering to a point that represents an absolute position and increase the number of clusters beyond the number of objects, the number of entries in the dictionary to be referred to increases. As the number of entries in the dictionary increases, the pointer to the dictionary must grow in size accordingly. For example, if the number of entries in the dictionary is 256, the pointer size can refer to the dictionary entry as at least 8 bits, whereas if the number of entries increases to 512, a 9-bit pointer is required.

In addition, the reason for separating the differential vector into length and angle before applying the clustering in the present invention is as follows.

First, in the differential vector extraction process, each vector is distributed at a similar angle to the road. In this case, when the clustering technique of grouping similar values into one value is applied, it is more advantageous in terms of location accuracy. Second, when the dictionary is formed, the number of possible values increases. If you create a dictionary with 10 entries by clustering in two-dimensional space, you can actually have 10 values. However, if you create two dictionaries with five entries each divided by two arguments, the actual representable value is 5 x 5. Therefore, it is separated in that more expression is possible with less dictionary. Finally, the lossy compression of map data requiring high positional accuracy results in a large number of clusters K. In this case, because the cost required for the compression process is increased, the length and the angle are separated to perform the clustering process.

FIG. 6 is a diagram illustrating a hardware configuration for implementing the present invention. When data about a geometric object is input through the input unit 100, the differential vector extracting unit 200 converts the geometric object into a starting point coordinate and a differential vector. The compression unit 300 separates the differential vectors into lengths and angles, and then applies a K-average clustering to produce a dictionary for each of lengths and angles. Compression is performed by converting a pointer to a representative value of a cluster to which the cluster belongs, and a starting point coordinate, a length dictionary, an angle dictionary, a length dictionary pointer, an angle dictionary pointer, etc. generated in this process are stored under the control of the controller 400. The starting point coordinate DB (510), length dictionary DB (520), angle dictionary DB (530), length dictionary pointer DB (540), angle dictionary pointer DB (550), etc. Each will be stored.

On the other hand, the data that has been compressed through the above process can be reconstructed into the original data through the reverse process of the compression process. To this end, the mobile device is provided with a restoring unit that performs the reverse process of the compression process, and receives the compressed data as described above, stores it in a storage space, and restores it through the restoring unit.

That is, the restoring unit of the mobile device obtains an entry from the storage space of the mobile device so that the pointer indicating the respective length and angle dictionary has the actual length and angle. This process yields an array of arrays of actual lengths and angles rather than arrays of pointers. When calculated using these lengths and angles, they are reconstructed into sets of differential vectors with relative position coordinates. By using the differential vector and the starting point of the object having absolute position coordinates of the spatial object, the original spatial object can be reconstructed.

Look at the experimental example of the present invention as described above.

The research area was selected around Yangcheon-gu and Gangseo-gu, Seoul (8.8km x 10.1km), and the data used in the experiment was extracted from features of building code from 1 / 1,000 numerical map of the study area. A total of 72016 polycon data, the total number of points is 566607, the experimental data as shown in FIG.

First, differential vectors for applying clustering techniques to the experimental data were extracted. The differential vector used in this study was extracted from the current point coordinates using the difference from the previous point coordinates, and in the case of the starting point of each object, the difference from the origin point was used.

The next step is to store the starting vector of each polygon data separately from the calculated differential vector. It can be seen that the differential vectors excluding the starting point are concentrated in the central portion of the differential vector of FIG. 8 by subtracting the number of polygons from the total number of points. If the data included a large area larger than the average length of the differential vector except for the starting point, clustering by the length and angle of the vector did not yield good results.

9 is an enlarged view of the differential vector excluding the starting point, and the obtained results are separated into lengths and angles, respectively, and are compressed as a process of obtaining two dictionaries and pointers by applying the K-average clustering technique and the dictionary-based compressor technique.

In the experiment, K-means clustering was performed using SPSS v10, a statistical package. The initial value K was clustered by increasing the distance and angle to 256, 512, and 1024. Fig. 10 shows the number of cases where the distance and angle generated in this way can be combined in advance.

In FIG. 10, the number of all cases that can have 256, 512, and 1024 dictionaries from the left as the number is 65536, 262144, and 1048576, respectively. This means that when the differential vector shown in FIG. 9 is approximated to a value closest to the dictionary of FIG. 10, the larger the dictionary, the smaller the error may be included.

FIG. 11 is a distribution of dictionaries that can be applied to experimental data using conventional research methods, and FIG. 11 represents a dictionary having 256, 512, and 1024 vectors from left. This is the difference in the number of vectors the dictionary can represent when a dictionary with the same amount of data is produced in two different ways.

In order to approximate each differential vector in advance, the compressed data was recovered and compared with the original to calculate the positional error, final data size, and compression rate due to data loss. Experiments were made by increasing the dictionary sizes to 65536, 262144, and 1048576, respectively. The size of the compressed data is the sum of the parts storing the starting point of the polygon, the two dictionaries representing the distance and angle of the differential vector, and the five parts of the two pointers to the dictionaries. The compression rate

Compression ratio (%) = ((original data size-compressed data size) / original data size) was calculated by the formula. The size of the original data is 9065712 bytes.

12 is a table for analyzing the results, and it can be seen that the position error level is significantly smaller in each of three experiments.

13, 14 and 15 are the results of experiments while increasing the dictionary size of the length and angle, respectively, Figure 13 is a dictionary size 256X256, Figure 14 is a dictionary size 512X512, Figure 15 is a dictionary size 1024X1024. In the case of FIG. 15, the data size was reduced to about 25% while maintaining accuracy.

The prior size of the existing study was adjusted to show results similar to those of the method presented in the present invention. In addition, experiments were made to include the starting point in advance so that two adjacent objects do not merge. The experimental results are shown in the table of FIG. In three experiments, experiments were carried out with increasing the size of the dictionary to 72272, 82528, and 73040. The dictionary is the result of K mean clustering using 72016 starting points and using 256, 512, and 1024 initial values, respectively. As can be seen from the tables of FIG. 12 and FIG. 16, it can be seen that the compression ratio of the present invention is high.

FIG. 17 shows the relationship between the compression rate and the positional error, and shows better positional accuracy when the compression rate is similar to that of the previous studies.

As described above, the present invention performs clustering by dividing the length and angle into two factors to cover various cases of the differential vector representing the relative position in one object, and to define the absolute position in the coordinate system. In the present invention, one point of the spatial object can be stored as it is without modification of the starting point, thereby improving the positional accuracy without reducing the compression rate. The data size is about 25% while minimizing the loss of the vector data. Can be lowered.

As described above, although described with reference to a preferred embodiment of the present invention, those skilled in the art various modifications of the present invention without departing from the spirit and scope of the invention described in the claims below Or it can be changed.

As described above, the vector data compression method and system using the K mean clustering according to the present invention divides the differential vector representing the relative position of each geometric object by length and angle, and clusters each by applying the K mean clustering technique. After the length and angle dictionaries are created, using the dictionary-based compression technique, each object is converted to a pointer to the representative value of the cluster to which the object belongs, using a dictionary produced by clustering to minimize the computation of vector data. Through this, it is possible to obtain an improved compression rate with a difference in location accuracy loss that is difficult to distinguish, thereby reducing the storage space occupied by the mobile device.

Claims (5)

In a method for compressing vector data in a mobile environment, Separating each geometric object into a sequence of starting point coordinates and differential vectors;  Separating the differential vector into length and angle; Performing a K-average clustering on the separated lengths and angles to produce a dictionary having an average value of the cluster as an entry; And Approximating each length and angle to convert to a pointer to a dictionary entry; Vector data compression method using the K mean clustering, characterized in that comprises a. The method of claim 1, wherein the differential vector is obtained using a difference between a corresponding coordinate of an individual coordinate of each geometric object and a previous coordinate. The method of claim 1, wherein the starting point coordinates representing the poetic point of the geometric object are not modified. In a system for compressing vector data in a mobile environment, A differential vector extracting unit that separates each geometric object into a sequence of starting point coordinates and differential vectors from data about each geometric object input through the input unit; The differential vectors extracted by the differential vector extracting unit are separated into lengths and angles, respectively, and a K mean clustering is applied to produce a dictionary for each of lengths and angles. A compression unit that converts and compresses a pointer to a value; A storage unit having respective databases storing starting point coordinates from the differential vector extracting unit, a length dictionary, an angle dictionary, a length dictionary pointer, an angle dictionary pointer, etc. from the compression unit; And A control unit controlling the respective units; Vector data compression system using the K mean clustering, characterized in that comprises a. The vector data compression system of claim 4, wherein the differential vector is obtained by using a difference between a corresponding coordinate of an individual coordinate of each geometric object and a previous coordinate.

Images