KR20210133611A - Demographic analysis system and operating method of the same - Google Patents

Abstract

The present invention relates to a demographic analysis system and an operation method thereof. The demographic analysis system according to an embodiment of the present invention includes an image extractor and an image analyzer. The image extractor extracts, from a captured image, tiles in which the number of vertexes overlapping with a target region is greater than the number of vertexes non-overlapping with the target region, and generates target images. The image analyzer extracts features respectively corresponding to the target images based on a trained neural network model, and generates a demographic result of the target region based on the extracted features.

Description

DEMOGRAPHIC ANALYSIS SYSTEM AND OPERATING METHOD OF THE SAME

The present invention relates to a demographic analysis system and an operating method thereof, and more particularly, to a satellite image-based demographic analysis system and an operating method thereof.

Satellite imagery acquired through remote sensing technology can provide useful information about human activity and topography without visiting the area. Since satellite imagery provides a wide range of geographic information, it can be utilized for various statistics. Furthermore, with the development of technology, the resolution of the satellite image is increasing. Accordingly, through the satellite image, demographics of a specific region, and furthermore, the economic size of the specific region can be analyzed. In addition, a change in a specific area over time may be analyzed through the satellite image.

Conventionally, statistical analysis using satellite images was performed manually. However, with the development of data processing technology, there is a demand for analyzing a satellite image through an analysis model implemented in an electronic device or system. In particular, there is a need for a method for efficiently processing and analyzing satellite images to improve processing speed and analysis accuracy.

The present invention can provide a demographic analysis system and an operating method thereof that can be applied to various terrains and can improve the accuracy of analysis results.

A demographic analysis system according to an embodiment of the present invention includes an image extractor and an image analyzer. The image extractor generates target images by extracting tiles from the captured image in which the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region. The image analyzer extracts features respectively corresponding to target images based on the learned neural network model, and generates demographic results of the target region based on the extracted features.

For example, the image extractor may generate target images by extracting tiles having three or more vertices overlapping the target area.

For example, when using a rectangular satellite image tile, the image analyzer calculates the cross entropy error based on the measured value of the result of analyzing the labeled image through the first convolutional neural network and the demographic result corresponding to the labeled image. Calculate, based on the result of analyzing the unlabeled image through the first convolutional neural network and the result of analyzing the unlabeled image through the second convolutional neural network, calculate the mean square error, cross-entropy error and average The analytic model may be trained by adjusting the first convolutional neural network based on the squared error. As an example, the image analyzer may extract features respectively corresponding to target images based on the adjusted first convolutional neural network.

For example, the image analyzer may classify each of the extracted features into any one of preset classes, and remove features classified as a target class from among the preset classes. For example, the image analyzer may classify the extracted features into an urban class, a rural class, and a non-inhabited class, and the target class may be a non-inhabited class.

For example, the image analyzer may reduce the dimension of the extracted features to a dimension determined based on the variance of the extracted features. As an example, the image analyzer may generate dimensionally reduced features by removing at least a portion of the extracted features and reducing dimensions of the extracted features, based on the classified class of each of the extracted features, and generating the dimensionally reduced features. Based on the mean, variance, number, and correlation of , a statistical expression having a fixed length may be generated, and a demographic result may be generated based on the statistical expression. For example, the demographic result may include at least one of population, age, education level, number of households, and economic size, average income, number of business owners, and industrialization degree for the target area.

A demographic analysis system according to an embodiment of the present invention includes an image extractor, a learner, a feature extractor, a dimension reducer, and a result generator. The image extractor generates target images that at least partially overlap the target area from the captured image. The learner learns an analytic model based on the labeled images and the number of unlabeled images greater than the labeled images. The feature extractor extracts features of target images based on the learned analysis model, classifies each of the extracted features into any one of preset classes, and removes features classified as a target class from among preset classes, Generate pruned features. The dimension reducer reduces the dimensionality of the pruned features to generate dimensionally reduced features. The result generator generates a demographic result of the target region based on the dimensionally reduced features.

For example, the image extractor may generate target images by extracting tiles in which the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region.

In one example, the learner generates a first result by analyzing labeled images, a first neural network that analyzes unlabeled images to generate a second result, a first result and a demographic result corresponding to the labeled images A first error calculator that calculates a cross entropy error based on the measured value of , a second neural network that analyzes unlabeled images based on a weight dependent on the exponential moving average of the first neural network to generate a third result; a second error calculator for calculating a mean square error based on the difference between the second result and the third result, and a weighted sum calculator for calculating a total loss by performing a weighted sum operation on the cross entropy error and the mean square error have. As an example, the weighted sum calculator calculates the total loss by adding the cross entropy error to the weighted mean square error, and the weight given to the mean square error may increase for a set number of learning times. As an example, the learner may adjust the weight of the first neural network until the total loss is within the reference range.

For example, the target class may be a non-human class. As an example, the dimensionality reducer may include a principal component analyzer that extracts principal components having a maximum variance from the pruned features and determines a reduced dimension of the pruned features based on the extracted principal components. In one example, the result generator is a statistical calculator that generates a statistical expression having a fixed length based on the mean, variance, number, and correlation of the dimensionally reduced features, and a demographic result based on a regression analysis of the statistical expression. It may include a regression analyzer that generates

The method of operating a demographic analysis system according to an embodiment of the present invention includes the steps of receiving an image, extracting tiles in which the number of vertices overlapping a target area in the image is greater than the number of vertices overlapping the target area, generating images, classifying features of each of the target images through a convolutional neural network into one of an urban class, a rural class, and a non-inhabited class, and based on the features classified into the urban class and the rural class generating demographic results.

In one example, the method includes receiving labeled images and unlabeled images, analyzing the labeled images through a convolutional neural network, and calculating a cross-entropy error of the convolutional neural network, the convolutional neural network and the first 2 analyzing unlabeled images through the convolutional neural network to calculate a mean squared error between the convolutional neural network and the second convolutional neural network, and based on the cross-entropy error and the mean squared error, the weight of the convolutional neural network It may further include the step of adjusting

As an example, generating the demographic result based on the features classified into the city class and the rural class may include reducing the dimensions of the features classified into the city class and the rural class, calculating statistics of the dimensionally reduced features, , generating a statistical representation having a fixed length, and calculating a demographic result of the target region based on the statistical representation.

A demographic analysis system and an operating method thereof according to an embodiment of the present invention may extract tiles corresponding to a target area to improve analysis accuracy and data processing speed of the target area. In addition, the demographic analysis system and the operating method thereof according to an embodiment of the present invention can improve the accuracy and reliability of the demographic results of the target area by using an analysis model optimized for generating the demographic results.

1 is a block diagram of a demographic analysis system according to an embodiment of the present invention.
FIG. 2 is a diagram for explaining an operation of the image extractor of FIG. 1 .
3 is an exemplary block diagram of the image analyzer of FIG. 1 .
FIG. 4 is an exemplary diagram of the learner of FIG. 3 .
5 is an exemplary diagram of the feature extractor of FIG. 3 .
6 is an exemplary diagram of the dimension reducer of FIG. 3 ;
7 is an exemplary diagram expressing features generated under the above-described process in a three-dimensional vector space.
FIG. 8 is an exemplary diagram of the result generator of FIG. 3 ;
9 is an exemplary block diagram of the demographic analysis system of FIG. 1 ;
10 is an exemplary flowchart of an operating method of the demographic analysis system described with reference to FIGS. 1 to 9 .

Hereinafter, embodiments of the present invention will be described clearly and in detail to the extent that those skilled in the art can easily practice the present invention.

Components described with reference to terms such as -part or unit, module, block, -or, -er, etc. used in the detailed description and functional blocks shown in the drawings These may be implemented in the form of software, hardware, or a combination thereof. Illustratively, the software may be machine code, firmware, embedded code, and application software. For example, the hardware may include an electrical circuit, an electronic circuit, a processor, a computer, an integrated circuit, integrated circuit cores, a pressure sensor, an inertial sensor, a microelectromechanical system (MEMS), a passive element, or a combination thereof. .

1 is a block diagram of a demographic analysis system according to an embodiment of the present invention. Referring to FIG. 1 , the demographic analysis system 100 includes an image extractor 110 and an image analyzer 120 . The demographic analysis system 100 may predict a demographic result PR of the target region by using the captured image. For example, the captured image may be an image captured by an artificial satellite, an aircraft, a drone, or various devices capable of capturing the ground at a high altitude. However, for the sake of simplicity, it is assumed that the captured image in the present specification is a satellite image (SI) captured by an artificial satellite.

The image extractor 110 may generate target images TI based on the satellite image SI. The image extractor 110 may receive the satellite image SI from the outside, and define a target area for demographic analysis from the satellite image SI. The image extractor 110 may extract target images TI corresponding to the target area. Each of the target images TI may be a tile having a predetermined shape. As an example, the tile may have a rectangular shape that is easy for convolutional neural network (CNN) analysis, but is not limited thereto.

The image extractor 110 may extract more tiles as the target images TI in which the number of vertices overlapping the target area is greater than the number of vertices non-overlapping the target area. For example, when the tile has a rectangular shape having four vertices, tiles having three or more vertices overlapping the target area may be extracted as the target images TI. Accordingly, images unnecessary for demographic analysis may be removed, and data for demographic analysis may be efficiently extracted even in complex terrain. Accordingly, the efficiency and accuracy of demographic analysis can be improved.

The image analyzer 120 may analyze each of the target images TI to generate a demographic result PR of the target region. The image analyzer 120 may extract features from the target images TI by using the learned neural network model. The image analyzer 120 may classify each of the target images TI based on the extracted features. For example, each of the target images TI may be divided into an area in which a person lives or an area in which a person does not live. For example, the image analyzer 120 may increase the analysis efficiency by removing the target images TI corresponding to an area not inhabited by a person. In addition, it may be analyzed whether an inhabited area is close to a rural area or an urban area. Through this analysis, a demographic result (PR) of the target area can be calculated. Details of the image analyzer 120 will be described later.

FIG. 2 is a diagram for explaining an operation of the image extractor of FIG. 1 . Referring to FIG. 2 , a satellite image SI in which the target area TA is captured is shown. The satellite image SI is generated by an artificial satellite and may be provided to the demographic analysis system 100 of FIG. 1 through wireless communication.

The image extractor 110 of FIG. 1 defines a target area TA from the satellite image SI. The target area TA refers to a geographic area for calculating demographic results. For example, the target area TA may be determined according to an administrative area. The image extractor 110 may extract target images TI corresponding to the target area TA.

As an example, the image extractor 110 of FIG. 1 extracts tiles in which the number of vertices existing inside the target area TA is greater than the number of vertices existing outside the target area TA as target images TI. can do. Accordingly, in FIG. 2 , tiles having three or more vertices in the target area TA may be extracted as target images TI. Accordingly, unnecessary images irrelevant to the administrative district to be analyzed may be removed. In addition, efficient demographic analysis is possible even for administrative districts with complex boundaries.

3 is an exemplary block diagram of the image analyzer of FIG. 1 . Referring to FIG. 3 , the image analyzer 120 includes a learner 121 , a feature extractor 122 , a dimension reducer 123 , and a result generator 124 .

The learner 121 may be implemented to learn a CNN-based analysis model for feature extraction of the target images TI. The learner 121 may use an analysis model by the Mean Teacher architecture. The learner 121 may receive a labeled image LI and an unlabeled image ULI for learning the analysis model. The labeled image LI contains randomly selected satellite images, and the satellite images are to be classified into three classes defined by three labels: urban, rural, and uninhabited. can The unlabeled image ULI may include a larger number of satellite images than the labeled image LI.

The labeled image LI may be analyzed by a first CNN (eg, a student network). The learner 121 may calculate a cross entropy error based on an analysis result of the first CNN and a correct answer (ie, ground truth). The unlabeled image (ULI) may be analyzed by a first CNN and a second CNN (eg, a teacher network). The weight of the second CNN may follow the exponential moving average of the first CNN. The learner 121 may calculate a mean squared error based on a difference between the analysis result of the first CNN and the analysis result of the second CNN. The learner 121 may calculate the final loss by adding the cross entropy error and the weighted mean square error. The learner 121 may fine-tune the analysis model based on the calculated final loss.

The feature extractor 122 may analyze the target images TI based on the analysis model MD fine-tuned by the learner 121 and extract features of the target images TI. Here, the fine-tuned analysis model MD may be the first CNN. The feature extractor 122 may classify each of the extracted features into one of classes such as urban, rural, and uninhabited. The feature extractor 122 may remove features classified as a non-human class. Practically, since the area in which a person lives is limited, the quality of information may be improved and the amount of information may be reduced according to the feature extraction (data pruning) of the feature extractor 122 . Accordingly, analysis efficiency and analysis speed in subsequent steps can be improved.

The dimension reducer 123 may reduce the dimension of the features PF pruned from the feature extractor 122 . The dimension reducer 123 may reduce the dimension of the pruned features PF through principal component analysis. The dimension reducer 123 may extract principal components having a maximum variance from the pruned features PF, and determine a dimension required to describe these principal components. The dimension reducer 123 may generate dimension-reduced features RPF to have the determined dimension.

The result generator 124 may generate a demographic result PR of the target region based on the dimensionally reduced features RPF. Result generator 124 may calculate statistics such as mean, standard deviation, number, and correlation (eg, Pearson's correlation coefficient) of dimensionally reduced features (RPF). The result generator 124 may generate a statistical expression having a fixed length based on the calculated statistics, and generate a demographic result (PR) through regression analysis on the generated expression. For example, the demographic result PR may include various demographic factors such as population, age, education level, number of households, and economic size for the target area. Furthermore, the demographic result PR may further include economic statistical factors such as average income, the number of business operators, and the degree of industrialization related to the target area.

FIG. 4 is an exemplary diagram of the learner of FIG. 3 . The learner 121 may use an analysis model by the Mean Teacher architecture. Referring to FIG. 4 , the learner 121 includes a first CNN (SN), a second CNN (TN), a cross entropy error calculator (CC) (hereinafter, a CEE calculator), and a mean square error calculator (MC) (hereinafter, MSE). calculator), and a weighted sum calculator (WC).

The first CNN (SN) may be a student network in the Mean Teacher architecture. The first CNN (SN) may analyze the labeled image (LI). The labeled image LI includes randomly selected satellite images. The labeled image LI may be assigned one of three labels: urban, rural, and uninhabited. The first CNN(SN) may classify the labeled image LI into one of urban, rural, and uninhabited classes.

The CEE calculator CC may calculate a cross entropy error for the result analyzed by the first CNN SN, that is, the first loss LL. The first loss LL may be calculated as in Equation 1.

Referring to Equation 1, g _i is defined as the correct answer class probability (ie, urban, rural, uninhabited classes), and f(d)i is the It is defined as a result of analyzing the labeled image (LI) with the first CNN (SN).

The second CNN (TN) may be a teacher network in the Mean Teacher architecture. The first CNN (SN) and the second CNN (TN) may analyze the unlabeled image (ULI). The unlabeled image ULI may include a larger number of satellite images than the labeled image LI. The weight of the second CNN may be determined according to the exponential moving average (EMA) of the first CNN. The analysis result of the first CNN (SN) and the analysis result of the second CNN (TN) may be provided to the MSE calculator (MC).

The MSE calculator MC may calculate a mean square error, that is, a second loss LU, based on the analysis result of the first CNN and the analysis result of the second CNN. The second loss LU may be calculated as in Equation (2).

Referring to Equation 2, f(d) is defined as a result of analyzing an unlabeled image (ULI) with a first CNN (SN), and t(d) is an unlabeled image (ULI) with a second CNN It is defined as the result of analysis with (TN).

The weighted sum calculator WC may calculate a total loss LT based on the first loss LL and the second loss LU. The weighted sum calculator WC may calculate the final loss LT by adding the first loss LL and the weighted second loss LU. As an example, the weight given to the second loss (LU) may increase (eg, from 0 to 12.5) during an initially set period (eg, 40 epochs). The first CNN (SN) is fine-tuned based on the total loss (LT). The learner 121 may iteratively fine-tune the first CNN SN so that the first loss LL, the second loss LU, and the total loss LT decrease within a reference range. Here, the reference range may be an acceptable range such that the accuracy of class classification of the first CNN (SN) can guarantee a specific accuracy of the demographic result. The fine-tuned first CNN (SN) is used to analyze the target images (TI) in the feature extractor 122 of FIG. 3 .

5 is an exemplary diagram of the feature extractor of FIG. 3 . The feature extractor 122 may use an analysis model learned by the Mean Teacher architecture. Referring to FIG. 5 , the feature extractor 122 may analyze target images TI based on the CNN (SN) fine-tuned by the learner 121 .

The CNN (SN) corresponds to the first CNN (SN) learned in FIG. 4 . The CNN (SN) may extract features of the target images (TI). CNN(SN) may classify each of the extracted features into one of classes such as urban, rural, and uninhabited. The feature extractor 122 may output pruned features PF by removing features classified as non-human classes. The pruned features PF may include n features v1, v2, ..., vn corresponding to cities or countryside. As an example, n may be the number of pruned target images, and each of the n features (v1, v2, ..., vn) may be expressed in an s dimension. As an example, s may be the number of pixels in the image.

6 is an exemplary diagram of the dimension reducer of FIG. 3 ; The dimension reducer 123 may reduce the dimension of the pruned features PF. Referring to FIG. 6 , the dimension reducer 123 may include a principal component analyzer (PCA).

The principal component analyzer PCA may reduce the dimension of the pruned features PF through principal component analysis. A principal component analyzer (PCA) may extract the principal components with maximum variance from the pruned features (PF) and determine the dimension k required to describe these principal components. Here, k is smaller than s, and may be, for example, one of values from 1 to 10. The principal component analyzer (PCA) may generate dimensionally reduced features (RPF) through linear orthogonal transformation of the pruned features (PF). The dimensionally reduced features RPF may include n features v1', v2', ..., vn'. Each of the n features v1', v2', ..., vn' may be expressed in k dimensions.

7 is an exemplary diagram expressing features generated under the above-described process in a three-dimensional vector space. Referring to FIG. 7 , features are distributed in a 3D vector space. As an example, the distributed features may be the dimensionally reduced features (RPF) described above. Each of the features corresponds to each of the above-described target images TI.

Referring to FIG. 7 , a distribution of features classified as a city and a distribution of features classified as a countryside may appear different from each other. For example, features corresponding to cities may have high variance along the y-axis and high mean along the z-axis. For example, features corresponding to rural areas may have a lower variance than features corresponding to cities. The features corresponding to the countryside may have a lower mean in the z-axis than the features corresponding to the city.

Target images TI1 to TI6 corresponding to each of the features are illustrated by way of example. The first target image TI1 and the second target image TI2 may be classified into a rural class. As the arrow corresponding to the countryside progresses, there may be a strong tendency to be classified as a rural area. That is, as the first target image TI1 progresses to the second target image TI2 , the degree of ruralization is strong and the population to be calculated later may be calculated to be low. The third to sixth target images TI3 , TI4 , TI5 , and TI6 may be classified into city classes. As the arrow corresponding to a city progresses, a characteristic with a strong tendency to be classified as a city may exist. That is, as the third target image TI3 progresses to the sixth target image TI6, the degree of urbanization is strong, and the population to be calculated later may be calculated to be high.

FIG. 8 is an exemplary diagram of the result generator of FIG. 3 ; The result generator 124 may generate the demographic result PR of the target region based on the dimensionally reduced features RPF generated through the above processes. Referring to FIG. 8 , the result generator 124 may include a statistical calculator 124_1 and a regression analyzer 124_2 .

The statistical calculator 124_1 may calculate statistical factors for the dimensionally reduced features RPF in order to generate a statistical representation Ri, which is data having a fixed length from a huge number of regions. As an example, the statistical calculator 124_1 calculates the average of the dimensionally reduced features (RPF), the standard deviation of the dimensionally reduced features (RPF), the number of the dimensionally reduced features (RPF), and the dimensionally reduced features ( RPF) can be calculated (eg, Pearson's correlation coefficient). The statistical calculator 124_1 may generate a statistical representation Ri having a fixed length by concatenating the calculated statistics. For example, the statistical calculator 124_1 may generate the statistical expression Ri by concatenating the calculated statistics using the Concat function.

The regression analyzer 124_2 may generate a demographic result PR of the target region through regression analysis on the statistical expression Ri. The demographic result (PR) may include various demographic factors such as population, age, education level, number of households, and economic size for a target area. For example, when there are many urbanized areas, the population, education level, number of households, economic size, etc. may appear high. For example, if there are many rural areas, the population, education level, number of households, economic size, etc. may appear low.

9 is an exemplary block diagram of the demographic analysis system of FIG. 1 ; Referring to FIG. 9 , the demographic analysis system 1000 may include a network interface 1100 , a processor 1200 , a working memory 1300 , a storage 1400 , and a bus 1500 .

The network interface 1100 is configured to communicate with external electronic devices. The network interface 1100 may receive a satellite image generated from an artificial satellite, and provide the received satellite image to the processor 1200 , the working memory 1300 , or the storage 1400 through the bus 1500 . The network interface 1100 may transmit a demographic result generated by analyzing a satellite image to an external device.

The processor 1200 may function as a central processing unit of the demographic analysis system 1000 . The processor 1200 may perform control operations and calculation operations required for data management, learning, and prediction of the demographic analysis system 1000 . For example, under the control of the processor 1200 , the network interface 1100 may receive a satellite image. Under the control of the processor 1200 , an analysis model may be learned, and a demographic result may be calculated using the learned analysis model. The processor 1200 may operate by utilizing the operation space of the working memory 1300 , and may read files for driving an operating system and executable files of applications from the storage 1400 . The processor 1200 may execute an operating system and various applications.

The working memory 1300 may store data and process codes to be processed or to be processed by the processor 1200 . For example, the working memory 1300 may include satellite images, information for extracting a target image, information for learning an analysis model, information for calculating demographic results, and information for building an analysis model. can be saved The working memory 1300 may be used as a main memory device of the demographic analysis system 1000 . The working memory 1300 may include a dynamic RAM (DRAM), a static RAM (SRAM), a phase-change RAM (PRAM), a magnetic RAM (MRAM), a ferroelectric RAM (FeRAM), and a resistive RAM (RRAM).

The image extractor 1310 and the image analyzer 1320 may be loaded into the working memory 1300 and executed. The image extractor 1310 and the image analyzer 1320 correspond to the image extractor 110 and the image analyzer 120 of FIG. 1 , respectively. The image extractor 1310 and the image analyzer 1320 may be part of an operation space of the working memory 1300 . In this case, the image extractor 1310 and the image analyzer 1320 may be implemented as firmware or software. For example, the firmware may be stored in the storage 1400 and loaded into the working memory 1300 when the firmware is executed. The processor 1200 may execute firmware loaded into the memory 1300 .

The image extractor 1310 may extract, as target images, more tiles from the satellite image in which the number of vertices overlapping the target area is greater than the number of vertices that do not overlap the target area. The image extractor 1310 may load a satellite image stored in the storage 1400 under the control of the processor 1200 and extract target images.

The image analyzer 1320 may analyze the target images to generate demographic results of the target area. A learning unit 1321 that learns an analysis model for analysis of a target image, class classification, and data pruning, a feature extraction unit 1322 that analyzes target images based on the learned analysis model, and the dimension of the extracted features The image analyzer 1320 may include a reduced dimension reduction unit 1323 and a result generator 1324 that generates demographic results of the target region based on the reduced dimensions. The learning unit 1321 , the feature extracting unit 1322 , the dimension reducing unit 1323 , and the result generating unit 1324 are the learner 121 , the feature extractor 122 , the dimension reducer 123 of FIG. 3 , respectively. and result generator 124 .

The storage 1400 may store data generated for long-term storage by the operating system or applications, a file for driving the operating system, or executable files of applications. For example, the storage 1400 may store files for execution of the image extractor 1310 and the image analyzer 1320 . The storage 1400 may be used as an auxiliary storage device of the demographic analysis system 1000 . The storage 1400 may include a flash memory, a phase-change RAM (PRAM), a magnetic RAM (MRAM), a ferroelectric RAM (FeRAM), a resistive RAM (RRAM), and the like.

Bus 1500 may provide a communication path between the components of demographic analysis system 1000 . The network interface 1100 , the processor 1200 , the working memory 1300 , and the storage 1400 may exchange data with each other through the bus 1500 . Bus 1500 may be configured to support various types of communication formats used in demographic analysis system 1000 .

10 is an exemplary flowchart of an operating method of the demographic analysis system described with reference to FIGS. 1 to 9 . The steps of FIG. 10 may be performed in the demographic analysis systems 100 and 1000 described with reference to FIGS. 1 to 9 . For convenience of description, FIG. 10 is described with reference to the reference numerals of FIGS. 1 and 3 .

In step S110, the demographic analysis system 100 learns the analysis model. The analysis model is used to extract features of the target images TI, and an analysis model by the Mean Teacher architecture may be used. The learner 121 of the image analyzer 120 analyzes the labeled image (LI) and the unlabeled image (ULI), calculates a cross entropy error and a mean square error, and refines the analysis model to reduce these errors. Can be adjusted.

In step S120 , the demographic analysis system 100 may receive a satellite image SI in order to calculate a demographic result of the target area. The image extractor 110 may extract target images TI overlapping the target area from the satellite image SI. For example, the image extractor 110 may extract more tiles as the target images TI in which the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region.

In operation S130 , the demographic analysis system 100 may extract features of the target images TI based on the analysis model learned in operation S110 . The feature extractor 122 may analyze the target images TI based on the analysis model MD fine-tuned by the learner 121 and extract features of the target images TI. The feature extractor 122 may classify each of the extracted features into one of classes such as urban, rural, and uninhabited. The feature extractor 122 may generate pruned features PF by removing features classified as non-human classes.

In step S140 , the demographic analysis system 100 may reduce the dimension of the extracted features. The dimension reducer 123 may reduce the dimension of the pruned features PF through principal component analysis.

In operation S150 , the demographic analysis system 100 may generate a demographic result PR of the target region based on the dimension-reduced features RPF. The result generator 124 may calculate the statistical factors of the dimensionally reduced features (RPF) to generate a statistical representation having a fixed length. The result generator 124 may generate a demographic result (PR) through regression analysis on the statistical expression. For example, the demographic result PR may include various demographic factors such as population, age, education level, number of households, and economic size for the target area.

The contents described above are specific examples for carrying out the present invention. The present invention will include not only the above-described embodiments, but also simple design changes or easily changeable embodiments. In addition, the present invention will also include techniques that can be easily modified and implemented in the future using the above-described embodiments.

100, 1000: demographic analysis system 110: image extractor
120: image analyzer 121: learner
122: feature extractor 123: dimension reducer
124: result calculator

Claims (20)

an image extractor for generating target images by extracting tiles from the captured image, wherein the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region; and
A demographic analysis system comprising an image analyzer for extracting features corresponding to the target images, respectively, based on a learned analysis model, and generating demographic results of the target region based on the extracted features. According to claim 1,
The image extractor,
A demographic analysis system for generating the target images by extracting tiles having three or more vertices overlapping the target area. According to claim 1,
The image analyzer,
Calculate a cross-entropy error based on a result of analyzing a labeled image through a first convolutional neural network and an actual value of a demographic result corresponding to the labeled image,
Calculate the mean square error based on the result of analyzing the unlabeled image through the first convolutional neural network and the result of analyzing the unlabeled image through the second convolutional neural network,
A demographic analysis system for learning the analysis model by adjusting the first convolutional neural network based on the cross-entropy error and the mean square error. 4. The method of claim 3,
The image analyzer,
A demographic analysis system for extracting features respectively corresponding to the target images based on the adjusted first convolutional neural network. According to claim 1,
The image analyzer,
A demographic analysis system for classifying each of the extracted features into any one of preset classes, and removing features classified as a target class from among the preset classes. 6. The method of claim 5,
The image analyzer,
A demographic analysis system for classifying the extracted features into an urban class, a rural class, and a non-inhabited class, wherein the target class is the non-inhabited class. According to claim 1,
The image analyzer,
A demographic analysis system for reducing a dimension of the extracted features to a dimension determined based on the variance of the extracted features. According to claim 1,
The image analyzer,
based on each classified class of the extracted features, removing at least some of the extracted features and reducing the dimensions of the extracted features to generate dimensionally reduced features,
generate a statistical representation having a fixed length based on the mean, variance, number, and correlation of the dimension-reduced features;
a demographic analysis system that generates the demographic result based on the statistical representation. According to claim 1,
The demographic result is a demographic analysis system including at least one of population, age, education level, number of households, economic size, average income, number of businesses, and degree of industrialization for the target area. an image extractor for generating target images that at least partially overlap a target area in the captured image;
a learner for learning an analysis model based on the labeled images and a greater number of unlabeled images than the labeled images;
Based on the learned analysis model, extracting features of the target images, classifying each of the extracted features into any one of preset classes, and removing features classified as a target class from among the preset classes , a feature extractor that generates pruned features;
a dimension reducer for reducing the dimensions of the pruned features to generate dimensionally reduced features; and
and a result generator for generating demographic results of the target region based on the reduced-dimensional features. 11. The method of claim 10,
The image extractor,
A demographic analysis system for generating the target images by extracting tiles in which the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region. 11. The method of claim 10,
The learner is
a first neural network that analyzes the labeled images to generate a first result, and analyzes the unlabeled images to generate a second result;
a first error calculator for calculating a cross entropy error based on the first result and actual values of demographic results corresponding to the labeled images;
a second neural network that analyzes the unlabeled images based on a weight dependent on the exponential moving average of the first neural network and generates a third result;
a second error calculator for calculating a mean square error based on a difference between the second result and the third result; and
and a weighted sum calculator for calculating a total loss by performing a weighted sum operation on the cross entropy error and the mean square error. 13. The method of claim 12,
The weighted sum calculator is
calculating the total loss by adding the cross entropy error to the weighted mean squared error;
The weight given to the mean squared error is increased for a set number of learning demographic analysis system. 13. The method of claim 12,
The learner is
A demographic analysis system for adjusting the weights of the first neural network until the total loss is within a reference range. 11. The method of claim 10,
The target class is a non-human class demographic analysis system. 11. The method of claim 10,
The dimension reducer,
and a principal component analyzer for extracting principal components having a maximum variance from the pruned features and determining a reduced dimension of the pruned features based on the extracted principal components. 11. The method of claim 10,
The result generator is
a statistical calculator for generating a statistical representation having a fixed length based on the mean, variance, number, and correlation of the dimension-reduced features; and
and a regression analyzer for generating the demographic result based on a regression analysis of the statistical expression. A method of operating a demographic analysis system, comprising:
receiving a photographed image;
generating target images by extracting tiles from the captured image in which the number of vertices overlapping the target region is greater than the number of vertices that do not overlap the target region;
classifying the features of each of the target images into one of an urban class, a rural class, and a non-inhabited class through a convolutional neural network; and
and generating demographic results based on the features classified into the city class and the rural class. 19. The method of claim 18,
receiving labeled images and unlabeled images;
calculating a cross-entropy error of the convolutional neural network by analyzing the labeled images through the convolutional neural network;
analyzing the unlabeled images through the convolutional neural network and the second convolutional neural network, and calculating a mean square error between the convolutional neural network and the second convolutional neural network; and
The method further comprising adjusting a weight of the convolutional neural network based on the cross entropy error and the mean square error. 19. The method of claim 18,
The step of generating a demographic result based on the characteristics classified into the city class and the rural class comprises:
reducing the dimension of the features classified into the city class and the rural class;
calculating statistics of the dimension-reduced features to generate a statistical representation having a fixed length; and
calculating the demographic result of the target area based on the statistical representation.

Images