JP2011039974A - Image search method and system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image search method and system, which have tolerance for a shooting angle of a query image and allow high-speed search with a small-scale system configuration. <P>SOLUTION: The image search system includes: a reference image extraction part 202 which extracts a reference image from a panoramic image of a town scape; a different-angle image generation part 203 which generates a different-angle image in which the angle of view of a subject is different from that of the reference image; a local feature quantity extraction part 204 which extracts local feature quantities from feature points of the reference image and the different-angle image; a corresponding point relation extraction part 205 which compares the respective feature points of the reference image with those of the different-angle image to extract a corresponding point relation having high degree of similarity between local feature quantities; and a filtering part 206 which narrows down feature points of the reference image to feature points having the corresponding point relation, and stores the narrowed-down feature points in a database 207. Local feature quantities extracted from respective feature points of the query image are compared with those extracted from the respective narrowed-down feature points of each reference image to obtain a reference image similar to the query image as a search result. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an image retrieval method and system for retrieving an image similar to a query image obtained by capturing a cityscape from a set of cityscape images that have been captured and stored in advance, and is particularly resistant to the imaging angle of a query image. The present invention relates to an image search method and system that enable high-speed search with a small system configuration.

  2. Description of the Related Art Navigation systems that guide users in real time using a mobile phone that can acquire GPS data have become widespread. However, in areas such as high-rise buildings where GPS signals are difficult to receive, position acquisition accuracy is unstable and errors are likely to occur. Therefore, there is a need for a system that can acquire accurate position information even in such areas.

  As a method of acquiring highly accurate position information in such an environment, image matching is performed between a cityscape image captured in the past to which position information is added and a captured image (query image) near the current user. A method for estimating position information is conceivable.

  In Patent Literature 1, a local feature amount in an image is extracted from a cityscape image and a photographed image taken in the past, and a cityscape image most similar to the photographed image is searched by comparing the local feature amounts of the images. A technique for estimating the position information added to the cityscape image as the current position of the user is disclosed.

  In Non-Patent Document 1, SIFT (Scale Invariant Feature Transform) is previously extracted as a local feature amount from both the query image and the search target image, and nearest neighbor search is performed based on the Euclidean distance L between the local feature amounts of each image. A technique is disclosed in which local feature quantities that are executed and have a short distance are made to correspond to each other, and a search target image having a large correspondence is finally used as a search result.

Japanese Patent Application No. 2009-049158

David G. Lowe, "Distinctive image features from scale-invariant keypoints" International Journal of Computer Vision, 60, 2 (2004), pp.91-110.

  Considering the capacity of the database, it is impractical to make the cityscape images to be searched continuous video images, but it will be a set of still images taken at regular intervals by an omnidirectional camera with the cityscape attached to a moving object. . Therefore, it is not possible to acquire all apparent photographing data for a certain subject (for example, a building), and the data becomes discrete. In order to accurately collate such a cityscape image database with an image (query image) taken by the user of the cityscape, a feature point highly resistant to changes in the shooting angle is required.

  On the other hand, in order to realize a search highly resistant to the shooting angle of the query image, a large number of feature points may be extracted from the cityscape image. However, storing all the local feature values obtained from a large number of feature points not only increases the capacity of the database, but also compares the local feature values extracted from the query image with a large number of local feature values during image search. Search time will be long because it has to be.

  An object of the present invention is to provide an image search method and system that solves all the technical problems described above, is resistant to the query image shooting angle, and enables high-speed search with a small system configuration. is there.

  In order to achieve the above object, the present invention is characterized in that the following measures are taken in an image retrieval system for retrieving an image similar to a query image from a set of cityscape images.

  (1) means for storing a large number of cityscape images, means for extracting a reference image from the cityscape images, means for generating different angle images in which at least a part of the reference image and the subject overlap and the angle at which the subject is viewed is different The means for extracting the local feature amount from each feature point of the reference image and the different angle image is compared with the feature point of the reference image and the different angle image, and the corresponding point relationship having a high similarity of the local feature amount is extracted. Means, filtering means for narrowing the feature points of the reference image to feature points in the corresponding point relationship, a database for storing the narrowed feature points of the reference image, and local features extracted from each feature point of the query image And means for comparing a quantity with a local feature quantity extracted from each narrowed-down feature point of each reference image, and using a reference image similar to a query image as a search result. That.

  (2) The database stores a large number of cityscape images in association with the position information, receives the query image from the user terminal, and returns the position information of the reference image similar to the query image to the user terminal. Means.

  According to the present invention, the following effects are achieved.

  (1) Since the feature points of the reference image to be searched are narrowed down only to the feature points that are resistant to the query image shooting angle, it is small without impairing the resistance to the query image shooting angle (geometric transformation). High-speed search is possible with a large-scale system configuration.

  (2) The reference image is stored in association with the position information, the image of the cityscape taken by the user terminal is used as a query image, a similar reference image is searched, and the position information is returned to the user terminal. As a result, the current position can be accurately and quickly guided to the user terminal.

It is the block diagram which showed the structure of the principal part of the position guidance system with which the image search device of this invention is applied. It is the figure which showed the relationship between a panoramic image and a reference | standard image. It is the figure which showed typically the provisional registration method of a corresponding point relationship. It is the figure which showed typically the corresponding registration method of a corresponding point relationship. It is the functional block diagram which showed the structure of the position guide part. It is the flowchart which showed operation | movement of one Embodiment of this invention. It is the flowchart which showed the temporary registration procedure of the corresponding point relationship. It is the flowchart which showed the main registration procedure of a corresponding point relationship. It is the figure which showed typically the procedure which narrows down the feature point of a reference | standard image to the feature point with high tolerance with respect to geometric transformation. It is the functional block diagram which showed the structure of the position guide part which concerns on 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a main part of a position guidance system to which an image search device of the present invention is applied.

  The position guidance system is mounted on a moving body such as a vehicle, and the imaging device 1 captures a cityscape with an omnidirectional camera. The local feature value is extracted by selecting a reference image from a panoramic image of the cityscape and registered. A similar image is selected from the reference images using a cityscape image (still image) photographed by the registration unit 2 and the user terminal 4 as a query image, and the position information of the most likely similar image is sent to the user terminal 4 for location guidance. The position guide unit 3 that responds as information has a main configuration.

  In the photographing apparatus 1, the omnidirectional camera 101 includes a large number of imaging units and shoots omnidirectional still images. The position detection unit 102 measures the current position (shooting position) of the shooting timing using, for example, a GPS signal. The direction detection unit 103 detects the direction of the omnidirectional camera 101, and detects the direction of each captured image based on the direction. Each photographed still image is associated with its photographing position and orientation and stored in the storage unit 104 one after another.

  In the local feature amount registration unit 2, the panoramic image generation unit 201 generates a panoramic image based on a large number of still images taken by the omnidirectional camera 101. The reference image extraction unit 202 extracts a reference image Iw from the panoramic image. The reference image Iw is each still image obtained by equally dividing the panoramic image, for example, as shown in FIG. 2, and the number of divisions does not necessarily match the number of image pickup units mounted on the omnidirectional camera 101. .

  The different angle image generation unit 203 generates a different angle image Iw ± θ obtained by photographing the subject of each reference image Iw by a predetermined angle θ in the left-right direction. Therefore, at least a part of the subject overlaps the reference image Iw and the different angle image Iw ± θ, and the angles at which the subject is viewed are different from each other. In this embodiment, the cityscape image is configured as an omnidirectional image, and images in a plurality of directions with different viewpoint directions can be acquired from each shooting point. There are images taken from different angles, and these images can be different angle images Iw ± θ. Such a different angle image Iw ± θ can also be obtained by cutting the reference image Iw from the panoramic image and projecting it so that the photographing angle is shifted.

  The local feature quantity extraction unit 204 extracts a local feature quantity from each feature point of each reference image Iw and its different angle image Iw ± θ. In the present embodiment, association of local feature amounts obtained from feature points is performed using SIFT. That is, for specifying the local region, feature point extraction based on the extreme value in the scale space by Difference of Gaussian (DoG) disclosed in Non-Patent Document 1 is used. As a result of the feature point extraction, the position of the feature point and the range of the local region are calculated, and a luminance gradient direction histogram is used as the feature descriptor of the local region.

  Such a direction histogram calculates the luminance gradient of each pixel in the feature region, weights it to generate a histogram, and rotates the feature region in that direction based on the direction of the most bin region (orientation) The direction histogram in the luminance direction is generated again, the histogram is further divided into blocks, the direction histogram is calculated in each block, and this is normalized and vectorized. In this embodiment, since the luminance direction in the block is divided into 8 directions and the target area is divided into 16 areas, one feature descriptor has 8 * 16 = 128 dimensions.

  As feature of these feature descriptors, local region features are generated, so it is resistant to occlusion, scales are determined for feature points, and the image size is unchanged, and orientation is set in the image plane based on the luminance gradient. Since it is performed, it is invariable to rotation with respect to the image plane. Furthermore, since edge components are used, it is resistant to luminance changes. Such feature point detection is performed on all pixels of the image, but even if a certain feature point takes an extreme value, it is excluded from the feature region if it is inappropriate as a feature point.

In the present embodiment, the local feature amount fw (i, j) of each reference image Iw (i) is expressed by the following equation (1). Where subscript i is the reference image identifier, subscript j is the feature point identifier, pw (i, j) is the position of the feature point expressed in homogeneous coordinates, ow (i, j) is the orientation of the feature point, and σw (i, j) is a scale at which a feature point is found, and dw (i, j) is a feature descriptor.

fw (i, j) = {pw (i, j), ow (i, j), σw (i, j), dw (i, j)} (1)

Similarly, the local feature amount fw ± θ (k, l) of each different angle image Iw ± θ is expressed by the following equation (2). However, the subscript k is a different angle image identifier, and the subscript j is a feature point identifier.

fw ± θ (k, l) = {pw ± θ (k, l), ow ± θ (k, l), σw ± θ (k, l), dw ± θ (k, l)} 2)

  The corresponding point relationship extraction unit 205 includes a temporary registration unit 205a and a main registration unit 205b, compares local feature amounts of each reference image Iw and its different angle image Iw ± θ, and selects a feature point pair having a higher similarity. Corresponding point relationship.

The temporary registration unit 205a performs a nearest neighbor search for each local feature amount of each different angle image Iw ± θ for each local feature amount of the reference image Iw, and calculates the similarity based on the distance between the local feature amounts. calculate. In the present embodiment, the similarity between the local feature amounts is represented by the Euclidean distance L between the feature descriptors dw (i, j) and dw (k, j) given by the following equation (3).

L = | dw (i, j) −dw ± θ (k, l) | (3)

  Then, for each feature point of the reference image Iw, the feature point having the closest Euclidean distance L and the feature point having the second closest distance L2 are searched from the different angle images Iw ± θ, and further each distance Only when L1 and L2 satisfy a predetermined first condition, a feature point pair that gives the distance L1 is provisionally registered as a corresponding point relationship.

  FIG. 3 is a diagram schematically showing a provisional registration method for corresponding points. In the illustrated example, attention is paid to a pair of the reference image Iw and one different angle image Iw + θ, and the local feature amount fw (1) of the reference image Iw and the local feature amount fw + of the different angle image Iw + θ. The Euclidean distance from θ (1) is L1, and the Euclidean distance from the local feature quantity fw + θ (2) is L2. Since the distance L1 is equal to or less than the predetermined reference distance Lref1 and the distance ratio L1 / L2 is equal to or less than the predetermined reference ratio Rref1 (for example, 0.8), the feature point pair that gives the Euclidean distance L1 is assumed as the corresponding point relationship. It is registered.

  In contrast, the Euclidean distance between the local feature fw (2) of the reference image Iw and the local feature fw + θ (3) of the different angle image Iw + θ is L1, and the local feature fw + θ (4) The Euclidean distance is L2, but the distance L1 is greater than the predetermined reference distance Lref1, or the distance ratio L1 / L2 is greater than the reference ratio Rref1 (for example, 0.8), so that the feature point that gives the distance L1 Even if it is a pair, it is not temporarily registered as a corresponding point relationship.

  The main registration unit 205b similarly performs a nearest neighbor search for local feature values obtained from each feature point of the reference image Iw with respect to the feature points on the different angle image Iw + θ side of all the corresponding correspondence points temporarily registered. If the feature points of the reference image Iw satisfying the main registration conditions equivalent to the temporary registration conditions are equal to the corresponding points on the reference image Iw side of the corresponding points, the corresponding points Fully register the relationship.

  FIG. 4 is a diagram schematically showing a corresponding registration method for corresponding points. In the illustrated example, the local feature quantity fw (1) of the reference image Iw and the local feature quantity fw + θ (1) of the different angle image Iw + θ, and the local feature quantity fw (2) of the reference image Iw and the different angle image Both the local feature quantity fw + θ (2) of Iw + θ are provisionally registered as corresponding point relationships. Then, local feature values fw + θ (1), fw + θ (2) obtained from corresponding points on the different angle image Iw + θ side of each corresponding point relationship and all local feature values fw (i) of the reference image Iw As a result of the nearest neighbor search, the feature point of the reference image Iw that satisfies the corresponding point relationship condition with the local feature amount fw + θ (1) corresponds to the reference image Iw side of the corresponding point relationship in the temporary registration Since the point coincides with the point, the corresponding point relationship is fully registered.

  On the other hand, the feature point of the reference image Iw that has a corresponding point relationship with the local feature amount fw + θ (2) does not match the corresponding point on the reference image Iw side of the corresponding point relationship in the temporary registration. Corresponding point relationship is not fully registered.

  Returning to FIG. 1, the filtering unit 206 does not have a corresponding point relationship with the feature points of both the different angle image Iw + θ and the other different angle image Iw−θ among all the feature points of the reference image Iw. The feature points of the reference image Iw (hereinafter, may be expressed as the feature points after filtering) are only deleted from the corresponding feature points of the different angle images Iw ± θ. Stored in the database 207. As a result, in the cityscape image database 207, the storage capacity of feature points that cannot be associated with corresponding points due to a difference in the shooting angle is not required, so that the database 207 can be reduced in size, and further, the query image and each reference Since the local feature amount when comparing images can be reduced, the search speed can be improved.

  The position guide unit 3 receives a cityscape image captured by the user terminal 4 and extracts local feature amounts from the feature points, and compares them with local feature amounts obtained from the filtered feature points of the reference image. Then, the corresponding point relationship is obtained, and the position information of the reference image Iw having the largest number of corresponding point relationships is returned to the user terminal 4 as position guidance information. The user terminal 4 recognizes the returned position information as its current position.

  FIG. 5 is a functional block diagram showing the configuration of the main part of the position guide unit 3. The position guide request receiving unit 301 receives a position guide request with a cityscape image Iq attached thereto from the user terminal 4. The local feature quantity extraction unit 302 extracts a local feature quantity fq (x) from the cityscape image Iq. Note that the subscript x is a feature point identifier of the cityscape image Iq.

  The corresponding point relationship extraction unit 303 includes a temporary registration unit 303a and a main registration unit 303b. Similar to the corresponding point relationship extraction unit 205, each local feature quantity fq (x) of the cityscape image Iq and each reference image Iw (k) Are compared with each local feature fw (k, l) of, and the corresponding point relationship is extracted. The similar image determination unit 304 determines the 1st to N best reference images Iw (k) having many corresponding points as similar images. The position information response unit 305 may respond to the user terminal with the position information of the 1 to N best reference image Iw (k) as the position guide information, or the 1 to N best reference image Iw (k). The shooting position of the cityscape image Iq may be estimated from the corresponding point relationship between the image and the cityscape image Iq, and the estimation result may be returned to the user terminal as position guidance information.

  6, 7, and 8 are flowcharts illustrating the operation of the present embodiment, and particularly illustrate the operations of the corresponding point relationship extraction unit 205 and the filtering unit 206. Here, as shown in FIG. 9 (a), the reference image Iw and its two different angle images Iw ± θ have already been generated, and the reference image Iw and each different angle image Iw ± θ In the following description, it is assumed that the extraction of the local feature amount from the feature points indicated in FIG.

  In step S1 of FIG. 6, one of the standard images Iw is selected as the current reference object, and in step S2, one of the different angle images Iw ± θ (here, the different angle image Iw + θ) is selected as the current reference object. Selected. In step S3, the temporary registration unit 205a compares the local feature obtained from each feature point of the reference image Iw with the local feature obtained from each feature point of the different angle image Iw + θ, and the local feature value is obtained. Temporary registration processing for temporarily registering matching or similar feature point pairs as corresponding point relationships is executed.

  FIG. 7 is a flowchart showing the procedure of the provisional registration process. In step S101, one of the local feature quantities fw (i, j) extracted from the current reference image Iw (i) is selected. In step S102, the similarity between each local feature quantity fwt (k, l) of the current different-angle image Iw + θ (k) and the current local feature quantity fw (i, j) of the reference image Iw is determined by each feature. Calculated as a distance (Euclidean distance) L between quantities. In step S103, the top two (L1, L2) feature points with the local Euclidean distance L close to each other are extracted. In step S104, it is determined whether or not the distance L1 is less than or equal to the reference distance Lef1 and the distance ratio (L1 / L2) is less than or equal to the reference ratio tref1. If both of these provisional registration conditions are satisfied, it is determined that the degree of similarity between the two is high, so the process proceeds to step S105, and the feature point pair that gives the distance L1 is provisionally registered as a corresponding point relationship.

  In step S106, it is determined whether or not the above processing has been completed for all local feature values fw (i, j) of the current reference image Iw (i). If not completed, the process returns to step 101, and the above processes are repeated while switching the local feature quantity fw (i, j) of the reference image Iw (i). Thereby, provisional registration of the corresponding point relationship between the current reference image Iw and the current different angle image Iw + θ is completed.

  Returning to the main flow of FIG. 6, in step S4, all the feature points on the different angle image Iw + θ side and the reference image Iw that have a corresponding point relationship in the main registration unit 205a with a line of sight opposite to that in step S3. The nearest neighbor search is performed on the local feature amount between the feature points. If the feature point of the reference image Iw having the highest local feature amount similarity matches the feature point on the reference image Iw side that is in the corresponding point relationship, the corresponding point relationship of the temporary registration is fully registered.

  FIG. 8 is a flowchart showing the procedure of the main registration process. In step S201, one of the temporarily registered corresponding points is selected. In step S202, the Euclidean distance L between the local feature amounts is calculated for the feature points on the different angle image Iw + θ side corresponding to the corresponding points and all the feature points of the reference image Iw. In step S203, the top two (L1, L2) feature point pairs with a short distance L are extracted.

  In step S204, it is determined whether or not the distance L1 is less than or equal to the reference distance Lef1 and the distance ratio (L1 / L2) is less than or equal to the reference ratio tref1. If all of these main registration conditions are satisfied, the process proceeds to step S205. In step S205, it is determined whether or not the feature point of the reference image Iw giving the distance L1 matches the corresponding point on the reference image Iw side of the corresponding point relationship. If they coincide with each other, the process proceeds to step S206, and the temporarily registered corresponding point relation is temporarily registered.

  In step S207, it is determined whether or not the above processing has been completed with respect to all the temporarily registered corresponding point relationships. If not completed, the process returns to step S201, and the above processes are repeated with respect to the remaining corresponding point relationships. Thus, the main registration of the corresponding point relationship between the current reference image Iw and the different angle image Iw + θ is completed. FIG. 9B is a diagram illustrating the registration result of the corresponding point relationship.

  Returning to the main flow of FIG. 6, in step S5, it is determined whether or not the above processing has been completed for all the different angle images Iw ± θ. Here, since the process has not been completed for the other different angle image Iw-θ, the process returns to step S2, and the above-described processes are repeated for the other different angle image Iw-θ.

  Thereafter, when the extraction of the corresponding point with the different angle image Iw + θ on the one side and the extraction of the corresponding point with the different angle image Iw-θ on the other side are completed for the current reference image Iw, the process proceeds to step S6, and the filtering unit 206 9 (c), both of one different angle image Iw + θ and the other different angle image Iw-θ are selected from the feature points first extracted from the current reference image Iw. Other feature points are discarded, leaving the feature points that are in the corresponding point relationship. As a result, as shown in FIG. 9D, the feature points of the reference image Iw are narrowed down to a small number.

  In step S <b> 7, the feature points narrowed down to the small number and their local feature amounts are registered in the database 207. In step S8, it is determined whether or not the above processing has been completed for all the reference images Iw. If not completed, the process returns to step S1, and the above-described processes are repeated while switching the reference image Iw.

  FIG. 10 is a functional block diagram showing the configuration of the position guide unit 3 according to the second embodiment of the present invention, and the same reference numerals as those described above represent the same or equivalent parts.

  In the present embodiment, when a position guidance request is received from the user terminal 4, the position estimation unit 301a that estimates the approximate current position of the user terminal 4 and the reference image Iw to be searched based on the estimation result of the current position are obtained. There is a feature in that a search range limiting unit 306 for limiting the area around the current position is provided, and the search range for position guidance is narrowed down in advance.

  The position estimation unit 301a is based on, for example, GPS signals transmitted from GPS satellites received by a GPS system installed in the user terminal 4 and insufficient for accurate positioning (for example, two), or The current location of the user terminal 4 is estimated based on the location of the base station that received the request message.

  According to the present embodiment, since the reference image Iw to be searched is narrowed down in advance, the search time can be shortened and the processing load of the search system can be reduced, and the search accuracy can be further improved.

  In each of the above embodiments, the two images adjacent to the left and right of the reference image Iw have been described as different angle images Iw ± θ. However, the present invention is not limited to this, and the angle Three or more different images are used as different angle images, and only the feature points that correspond to the feature points of all the different angle images Iw ± θ among the feature points first extracted from the reference image Iw are filtered. You may make it register as a later feature point.

  Furthermore, in the above-described embodiment, among the feature points extracted first from the reference image Iw, only the feature points that correspond to the feature points of all the different angle images Iw ± θ are the feature points after the filtering. Although described as being registered, the present invention is not limited to this, and at least some of the feature points of the different-angle image Iw ± θ are associated with the corresponding points after the filtering. It may be registered as a feature point.

  DESCRIPTION OF SYMBOLS 1 ... Imaging device, 2 ... Local feature-value registration part, 3 ... Position guidance part, 4 ... User terminal, 101 ... Omnidirectional camera, 102 ... Position detection part, 103 ... Direction detection part, 201 ... Panorama image generation part, 202 Reference image extraction unit, 203 ... Different angle image generation unit, 204 ... Local feature amount extraction unit, 205 ... Corresponding point relationship extraction unit, 206 ... Filtering unit, 207 ... Database

Claims (9)

In an image search system that searches an image similar to a query image from a set of cityscape images,
A means of storing a large number of cityscape images;
Means for extracting a reference image from the cityscape image;
Means for generating a different angle image in which at least a part of the reference image and the subject overlap and the angle at which the subject is viewed is different;
Means for extracting a local feature amount from each feature point of the reference image and the different angle image;
Means for comparing each feature point of the reference image and its different angle image and extracting a corresponding point relationship having a high similarity of local feature amounts;
Filtering means for narrowing down feature points of a reference image to feature points in the corresponding point relationship;
A database for storing the narrowed feature points of the reference image;
Means for comparing a local feature amount extracted from each feature point of a query image with a local feature amount extracted from each of the narrowed-down feature points of each reference image, and using a reference image similar to the query image as a search result An image search system comprising: The database stores a number of cityscape images associated with the location information,
Means for receiving the query image from a user terminal;
The image search system according to claim 1, further comprising means for responding to the user terminal with position information corresponding to a reference image similar to the query image. Means for estimating the current position of the user terminal;
The image search system according to claim 2, further comprising means for previously narrowing down a reference image to be searched based on the estimation result. The means for generating the different angle image generates a plurality of different angle images with different angles for viewing the subject for each reference image,
4. The image search system according to claim 1, wherein the means for extracting the corresponding point relationship extracts a corresponding point relationship with the reference image for each different angle image.   5. The image search system according to claim 4, wherein the filtering means narrows down the feature points of the reference image to feature points that correspond to the feature points of all the different angle images.   The image search system according to claim 4, wherein the filtering unit narrows down feature points of the reference image to feature points that correspond to the feature points of some different-angle images.   The image retrieval system according to claim 1, wherein the reference image is extracted from a panoramic image of a cityscape. In an image search method for searching an image similar to a query image from a set of cityscape images,
A procedure for extracting a reference image from a number of cityscape images stored in advance;
A procedure for generating a different angle image in which at least a part of the reference image and the subject overlap and the angle at which the subject is viewed is different;
A procedure for extracting a local feature amount from each feature point of the reference image and the different angle image;
Comparing each feature point of the reference image and its different angle image, and extracting a corresponding point relationship with a high similarity of local feature amounts;
A procedure for narrowing down feature points of a reference image to feature points in the corresponding point relationship;
Storing the narrowed feature points of the reference image in a database;
A procedure for comparing a local feature amount extracted from each feature point of a query image with a local feature amount extracted from each of the narrowed-down feature points of each reference image, and using a reference image similar to the query image as a search result An image search method comprising: In the database, many cityscape images are associated with the location information,
Receiving the query image from a user terminal;
The image search method according to claim 8, further comprising a step of responding to the user terminal with position information corresponding to a reference image similar to the query image.