JP2017010468A - System and method for searching for an object in a captured image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a system and the like reduce a data volume of local characteristics to be transferred via a network without reducing retrieval accuracy upon retrieving objects reflected in imaging pictures.SOLUTION: An object retrieval server is configured to preliminarily store reference pictures having a subject reflected, and a terminal is configured to transmit question information including a plurality of characteristic vectors in a question picture having the subject reflected to an object retrieval server. Then, the object retrieval server is configured to calculate a degree of similarity between the plurality of characteristic vectors of the question information and each reference pictures, and transmit a determination result of the degree of similarity between the question information and a plurality of reference pictures to the terminal. Next, the terminal is configured to transmit geometry information including coordinate information on the question picture in accordance with the received determination result of the degree of similarity to the object retrieval server. Next, the object retrieval server is configure to retrieve one or more reference pictures similar by more than a prescribed threshold to the coordinate information on the geometry information in the plurality of retrieved reference pictures, and reply a reference picture ID of each reference picture to the terminal.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a technique for searching for an object appearing in a captured image.

  Conventionally, there is a technique for searching for an object using local features extracted from an image (for example, see Non-Patent Document 1). According to this technique, an object is photographed by a camera, and a local feature vector is extracted from a photographed image (question (query) image) in which the object is photographed. Then, the local feature vector of the question image is compared with the local feature vector of the image (reference (reference) image) in which the object registered in advance is searched, and the object is searched. Further, the search accuracy is improved by executing geometric verification between the reference images with the highest N cases and the query images. Then, the number of feature point correspondence groups (inliers) that satisfy the conditions in the geometric verification is set as a new inter-image similarity.

  On the other hand, there is a case where it is desired to determine whether or not the same object is included in both the reference image and the question image using a threshold value based on the similarity. This is because the higher the threshold of similarity between images, the higher the recognition precision.

  FIG. 1 is a functional configuration diagram of a terminal and an object search server in the prior art.

  FIG. 1 shows a system for searching for an object appearing in a captured image using a mobile terminal with a camera. The portable terminal extracts a local feature vector from the captured image and quantizes the local feature vector. On the other hand, the object search server also extracts a large number of reference images or local feature vectors, quantizes the local feature vectors, and stores the quantized vectors in advance for each image ID. The portable terminal transmits the question information including the quantized vector of the captured image to the object search server via the network. The object search server searches for a reference image most similar to the question information, and returns the reference image ID to the portable terminal. Thereby, the portable terminal can recognize that the object shown in the captured image is based on the reference image ID.

  In order to execute such a service in real time, the size of the question information transmitted from the terminal to the search server via the network is important. The larger the data size, the longer the communication time and the real-time property is impaired. On the other hand, there is a technique for reducing the size of data to be transmitted by reducing the size of the local feature vector used for search (see, for example, Patent Document 1).

WO2013033622A1

Philbin. J, Chum. O, Isard. M, Sivic. J, and Zisserman. A, “Object retrieval with large vocabularies and fast spatial matching,” In Computer Vision and Pattern Recognition, pp. 1-8, 2007.

  According to the technique described in Patent Document 1, the size of data transferred via a network is reduced by reducing the size of the local feature vector. In general, the reduction of local feature vectors is thought to lead to a decrease in object search accuracy. On the other hand, can the inventors of the present application reduce the data size transferred through the network without reducing the object search accuracy? I thought. Is it possible to reduce the data size for “coordinate information” of local features? I thought.

  Therefore, the present invention provides a system and method capable of reducing the data size of local features transferred via a network without reducing the object search accuracy when searching for an object appearing in a captured image. The purpose is to do.

According to the present invention, there is provided an object search method for a system having a terminal and an object search server,
The object search server stores in advance a reference image in which an object is shown,
A first step in which a terminal transmits question information including a plurality of feature vectors in a question image in which an object is captured to an object search server;
A second step in which the object search server calculates a similarity between a plurality of feature vectors of question information and each reference image;
A third step in which the object search server returns a similarity determination result between the question information and the plurality of reference images to the terminal;
A fourth step in which the terminal selects geometric information including coordinate information of the question image according to the received similarity determination result;
A fifth step in which the terminal transmits the selected geometric information to the object search server;
A sixth step in which the object search server searches for one or more reference images similar to the coordinate information of the geometric information at a predetermined threshold value or more among the plurality of reference images searched in the second step;
The object search server includes a seventh step of returning the reference image ID of each reference image searched in the fifth step to the terminal.

According to another embodiment of the object search method of the present invention,
For the first step, the terminal
Extracting a local feature including a feature vector and coordinate information from a query image showing an object;
Quantizing the feature vector into a quantization vector;
Transmitting question information including a plurality of quantization vectors to an object search server,
For the second step, the object search server
It is also preferable to calculate the degree of similarity between the plurality of quantization vectors of the question information and each reference image.

According to another embodiment of the object search method of the present invention,
The quantization vector included in the question information is also preferably a Visual Word identifier.

According to another embodiment of the object search method of the present invention,
The question information preferably further includes a direction and / or a scale in association with the Visual Word identifier.

According to another embodiment of the object search method of the present invention,
For the third step, the object search server returns one of the following three commands as a similarity determination result to the terminal,
yes1: There is one or more reference images whose similarity is greater than or equal to the threshold TH1
yes2: There is one or more reference images whose similarity is less than the threshold TH1 and greater than or equal to the threshold TH2.
no: all similarities are less than TH2 threshold TH1: index of similarity with very high possibility of successful search threshold TH2: index of similarity with high possibility of successful search For the fourth step, the terminal The geometric information to be transmitted to the object search server is controlled as follows according to the similarity determination result
yes1: Geometric information includes only coordinate information
yes2: Geometric information includes coordinate information and feature vectors
no: It is also preferable to end the processing as a search failure.

According to another embodiment of the object search method of the present invention,
For the third step, the object search server returns one of four or more commands to the terminal as a similarity determination result with three or more threshold values,
Regarding the fourth step, it is also preferable that the terminal includes a longer feature vector (only upper bits of the feature vector-> total bits) in the geometric information as the threshold value based on the similarity determination result is lower.

According to another embodiment of the object search method of the present invention,
For the sixth step, the object search server
A feature point correspondence set satisfying a predetermined geometric condition is calculated between the coordinate information of the geometric information and the coordinate information of each reference image,
It is also preferable to search for a reference image in which the number of feature point correspondence groups is the maximum or a predetermined number or more.

According to the present invention, a system having a terminal and an object search server,
The terminal
Question information transmission means for transmitting question information including a plurality of feature vectors in a question image in which an object is captured to an object search server;
Geometric information selection means for selecting geometric information including coordinate information of the query image according to the received similarity determination result;
Geometric information transmission means for transmitting the selected geometric information to the object search server,
The object search server stores in advance a reference image in which an object is shown,
Similarity calculating means for calculating the similarity between the plurality of feature vectors of the question information and each reference image;
Similarity determination transmission means for returning the similarity determination result between the question information and the plurality of reference images to the terminal;
Geometric verification means for searching for one or more reference images that are similar to the coordinate information of the geometric information among a plurality of reference images determined to be similar based on the similarity determination result;
Search result transmission means for returning the reference image ID of each reference image searched by the geometric verification means to the terminal.

  According to the system and method of the present invention, when searching for an object shown in a captured image, the data size of local features transferred via a network can be reduced without reducing the object search accuracy.

It is a functional block diagram of the terminal and object search server in a prior art. It is a function block diagram of the terminal and object search server in this invention. It is a sequence diagram between the terminal and object search server in this invention. It is explanatory drawing showing a transposition index. It is explanatory drawing showing matching between a question image and a reference image. It is explanatory drawing showing the similarity of VWID (quantization vector) between a question image and a reference image. It is explanatory drawing showing the geometric feature point corresponding | compatible group between a question image and a reference image. It is another system block diagram in this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a functional configuration diagram of the terminal and the object search server in the present invention.
FIG. 3 is a sequence diagram between the terminal and the object search server in the present invention.

  According to the present invention, an object appearing in a captured image can be searched by a system having the terminal 1 and the object search server 2. The terminal 1 is, for example, a smartphone or a mobile phone, and generally has a camera. Captured images in which various objects are captured can be acquired in response to a user operation on the camera. On the other hand, the object search server 2 stores in advance a reference image showing an object. And the terminal 1 can identify the object reflected in the picked-up image by communicating with the object search server 2 via a network.

According to FIG. 2, the terminal 1 includes a local feature extraction unit 11, a quantization unit 12, a geometric information storage unit 13, a question information transmission unit 14, a geometric information selection unit 15, and a geometric information transmission unit 16. Have These functional components are realized by executing a program that causes a computer installed in the terminal to function.
Further, the object search server 2 communicating with the terminal 1 includes a local feature extraction unit 21, a quantization unit 22, a geometric information storage unit 23, a transposed index storage unit 24, a similarity calculation unit 25, and a similarity determination. A transmission unit 26, a geometric verification unit 27, and a search result transmission unit 28 are included. These functional components are realized by executing a program that causes a computer mounted on the server to function.

  In the following, each functional component of FIG. 2 will be described along the sequence of FIG.

  According to the prior art, the search for an object shown in a question image is executed in a one-step sequence as shown in FIG. In contrast, the present invention is executed in a two-stage sequence. In the first stage, a plurality of reference images that are similar in feature vector to the question image are searched, and in the second stage, a geometrically similar reference image is further searched from among the reference images.

<S01> [Terminal 1: Local feature extraction unit 11]
The local feature extraction unit 11 inputs an image captured by a camera and extracts a local feature in the image. A local feature represents a feature point position by, for example, the following elements.
Feature vector: f (eg 4 bytes)
Coordinate: p = (x, y) (for example, 4 bytes (x (2 bytes) y (2 bytes))
Direction: θ
Scale: s
The local feature extracted by the present invention includes at least a feature vector and coordinate information.

  Usually, a local feature group detected from one image is stored on an array, and a subscript of the array is identified as a “local feature ID”. Further, the feature vector f of the local feature is output to the quantization unit 12.

  For example, SIFT, SURF (Speeded Up Robust Features), or ORB (Oriented FAST and Rotated BRIEF) is used as the feature vector extraction algorithm used for object recognition. For example, in the case of SIFT, a set of 128-dimensional feature vectors is extracted from one image. SIFT is a technique for analyzing a characteristic local region using a scale space and describing a feature vector that is invariant to scale change and rotation. On the other hand, in the case of SURF, higher-speed processing is possible than SIFT, and a set of 64-dimensional feature vectors is extracted from one image. According to the ORB, a set of 256-bit binary local feature vectors is extracted from one content.

<S02> [Terminal 1: Quantization unit 12]
The quantization unit 12 receives the feature vector f from the local feature extraction unit 11. The quantization unit 12 stores a plurality of quantization vectors called Visual Word (VW) in advance. These quantization vectors are stored as a code book. Then, the quantization unit 12 quantizes the VW ID (VWID) closest to (similar to) the feature vector f. Here, the nearest VWID may be one, or the top N items having the shortest distance. A VWID is extracted for each feature vector f, and the VWID group is output to the question information transmission unit 14 as a quantization result.

<S03> [Terminal 1: Geometric Information Storage Unit 13]
The geometric information storage unit 13 inputs a local feature group from the local feature extraction unit 11. Then, the geometric information storage unit 13 stores the geometric information in a list structure in the order of the local feature IDs. The geometric information is a combination of coordinate information p and matching information. The matching information may be a VWID or a feature vector (the whole feature vector, a part of the feature vector, or an encoded feature vector). Note that the geometric information storage unit 13 preferably discards the geometric information when a predetermined time has elapsed after the storage of the geometric information.

[Object Search Server 2: Local Feature Extraction Unit 21]
The local feature extraction unit 21 inputs a large number of reference images in which an object is photographed, and extracts local features in each reference image. Its function is exactly the same as the local feature extraction unit 11 of the terminal 1 described above. The extracted feature vector is output to the quantization unit 22.

[Object Search Server 2: Quantization Unit 22]
The quantization unit 22 receives the feature vector f from the local feature extraction unit 21 and quantizes it into a VWID using a codebook. Its function is exactly the same as that of the quantization unit 12 of the terminal 1 described above. A group of VWIDs as a quantization result is output to the geometric information storage unit 23 and the transposed index storage unit 24.

[Object Search Server 2: Geometric Information Storage Unit 23]
The geometric information storage unit 23 inputs a local feature group from the local feature extraction unit 21. Its function is exactly the same as the geometric information storage unit 23 of the terminal 1 described above.

[Object Search Server 2: Transposed Index Storage Unit 24]
The transposed index storage unit 24 stores a reference image ID and local feature information in a list structure in association with each VWID (word ID).

  FIG. 4 is an explanatory diagram showing an inverted index.

  According to FIG. 4, the reference image ID and the local feature information are associated with each VWID. The local feature information includes a feature vector f, coordinate information p, direction θ, and scale s.

  FIG. 5 is an explanatory diagram showing matching between a question image and a reference image.

FIG. 5A shows that the viewpoint of the question image is different from the reference image. In addition, the object shown in the question image is geometrically transformed with respect to the object shown in the reference image.
FIG. 5B shows local features extracted from the question image and the reference image. The local features include at least a feature vector f and coordinate information p, and may further include a direction θ and a scale s.
FIG. 5C shows matching between feature points extracted from the question image and the reference image. The greater the number of matching feature point pairs, the higher the geometric similarity.

<S1> [Terminal 1: Question information transmitter 14]
The question information transmission unit 14 transmits the question information including a plurality of quantization vectors (feature vectors) in the question image in which the object is captured to the object search server 2. Here, the quantization vector is preferably a VWID (Visual Word identifier). The question information is generated by arranging the VWIDs output from the quantization unit 12 in the order of local feature IDs. That is, the question information is a group of VWIDs. When using Hamming embedding or Weak geometry consistency, the question information may further include a direction θ and / or a scale s in association with the Visual Word identifier.

According to the prior art, the search for an object appearing in a query image is executed in a one-step sequence as shown in FIG. 1, whereas the present invention is executed in a two-step sequence.
First stage: search for a plurality of reference images similar to the VWID group of the question image
(Question information does not include coordinate information. Calculate similarity by voting)
Second stage: Among the retrieved reference images, a reference image similar to the geometric information of the question information is selected.
Search further
(Perform geometric verification only when the similarity is greater than or equal to the threshold at the first stage)

For example, according to the prior art, when the question information is feature vector + coordinate information, 8 bytes are required. On the other hand, according to the present invention, in the first stage, since the question information can be only VWID, only half of 4 bytes are required. That is, the data size of the question information can be reduced by 50%.
When it is determined that there is no similar reference image by executing the first-stage sequence, it is not necessary to execute the second-stage sequence. That is, as the object search service continues, the data size transmitted / received via the network inevitably decreases.

<S2> [Server 2: Similarity Calculation Unit 25]
The similarity calculation unit 25 uses the transposed index storage unit 24 to calculate “similarity” between a plurality of quantization vectors of question information and each reference image. Specifically, the VWID group of many reference images is compared with the VWID group included in the received question information, and the similarity is calculated.

  FIG. 6 is an explanatory diagram showing the similarity of the VWID (quantization vector) between the question image and the reference image.

  Here, a score is voted for the reference image ID included in the list of transposed indexes corresponding to each VWID included in the question information (see FIG. 4). The score for voting may be a fixed size or some weighting. The larger the voted score, the higher the similarity between the question image and the reference image.

  Since the transposed index has a difference in the total number of VWs for each reference image, it is difficult to compare the sizes of the voted scores. Therefore, it is also preferable to normalize by dividing the similarity score for each reference image by the Lp norm of the VW histogram of each reference image.

It is assumed that VWIDi is set as follows.
xi: Number of assigned local features of the reference image n: Number of VWs prepared in advance Here, the VW histogram of the reference image is expressed as follows.
Vector x ^→ = (x ₁ , x ₂ ,..., X _n )
At this time, for example, the norm of the VW histogram of the reference image is expressed as follows.
L1 norm: Σ _{i = 1} ⁿ | x _i |
: Localization quantized to each VW in the VW histogram of the reference image
Total number of features L2 norm: √ (Σ _{i = 1} ⁿ x _i ² )
: Square root of the sum of squares of the number of local features quantized for each VW Normalization of similarity by multiplying the reciprocal of the L1 norm, the reciprocal of the L2 norm, or other normalization coefficient for each reference image ID Turn into. Finally, similarities corresponding to all reference image IDs are calculated.

<S3> [Object Search Server 2: Similarity Determination Transmitter 26]
The similarity determination transmission unit 26 returns the similarity determination result between the question information and the plurality of reference images to the terminal 1. The similarity determination result is determined with a threshold for the similarity of all reference image IDs with respect to the question information. For example, the similarity determination result is represented by three commands for two threshold values as follows.
(command)
yes1: There is one or more reference images whose similarity is greater than or equal to the threshold TH1
yes2: There is one or more reference images whose similarity is less than the threshold TH1 and greater than or equal to the threshold TH2.
no: all similarities are less than TH2
Threshold TH1: Similarity index with very high possibility of successful search
Threshold TH2: Similarity index with a relatively high probability of successful search

In the case of the commands “yes1” and “yes2”, the similarity determination transmission unit 26 transmits to the terminal 1 the reference image ID that satisfies the determination condition together with the command.
Further, in the case of the command “no”, the similarity determination transmission unit 26 regards that the search has failed and transmits only the command to the terminal 1. As a result, the terminal 1 can end as “search failure” without executing the second-stage sequence.

  Note that, as a similarity determination result, any of four or more commands may be returned to the terminal using three or more threshold values. Further, the number of reference image IDs may be predetermined N in descending order of similarity, or may be all.

<S4> [Terminal 1: Geometric Information Selection Unit 15]
The geometric information selection unit 15 receives the similarity determination result from the object search server 2. Then, in accordance with the similarity determination result, geometric information including the coordinate information of the question image is selected. Specifically, the geometric information transmitted to the object search server 2 is controlled as follows.
(Receive command)
yes1: Geometric information includes only coordinate information
yes2: Geometric information includes coordinate information and feature vectors
no: Terminates processing as a search failure

Here, it is also preferable that the geometric information selection unit 15 includes a longer feature vector (only upper bits of the feature vector-> whole bits) in the geometric information as the threshold value based on the similarity determination result is lower.
(Receive command)
yes1 (the judgment threshold is high): Geometric information includes only coordinate information
yes2: Geometric information includes coordinate information and part of feature vector
yes3 (decision threshold is low): Geometric information includes all coordinate information and feature vectors
no: Discards the stored geometric information after time t

<S5> [Terminal 1: Geometric Information Transmitter 16]
The geometric information transmission unit 16 transmits geometric information including the coordinate information of the question image to the object search server 2.

<S6> [Object Search Server 2: Geometric Verification Unit 27]
The geometric verification unit 27 searches for one or more reference images that are similar to the coordinate information of the geometric information by a predetermined threshold value or more among the plurality of reference images determined to be similar based on the similarity determination result (FIGS. 4 and 4). 5 (c)). Specifically, a feature point correspondence group (inlier) that satisfies a predetermined geometric condition is calculated between the coordinate information of the geometric information and the coordinate information of each reference image. Then, a reference image having the maximum number of feature point correspondence groups or a predetermined number or more is searched.

First, a feature point correspondence set in which coordinate information of local features quantized to the same VWID is matched by coordinate matching is calculated. For example, the closest feature vectors f may be matched.
Next, using a robust estimation method such as a RANSAC (RANdom SAmple Consensus) algorithm, a homography matrix from one coordinate system to the other coordinate system is calculated for the matched feature point correspondence group.
Next, the coordinates of one feature point are projected using the Homography matrix, and if the distance from the corresponding coordinate of the other feature point is within L, the feature point correspondence group is determined to be inlier.
Only reference images in which the number of inliers is equal to or greater than a predetermined threshold are determined as “search success”. However, if there is no reference image in which the number of inliers is equal to or greater than a predetermined threshold, it is determined as “search failure”.
Note that a matrix based on a geometric transformation such as an Affine matrix or a basic matrix may be used instead of the Homography matrix.

  FIG. 7 is an explanatory diagram illustrating a geometric feature point correspondence group between a question image and a reference image.

FIG. 7A shows a feature point correspondence set between a question image and a reference image.
FIG. 7B shows the correspondence between inlier and outlier images based on the Homography matrix. outlier is represented by a broken line.

<S7> [Object Search Server 2: Search Result Transmitter 28]
The search result transmission unit 28 returns the reference image ID of each reference image searched by the geometric verification unit 27 to the terminal 1.

  FIG. 8 is another system configuration diagram in the present invention.

  According to FIG. 8, the object search server of FIG. 2 is arranged separately as a similarity determination server and a geometric verification server. By constructing such separate servers, the processing load on each server can be reduced.

In this case, for S3 (similarity determination transmission unit of the object search server), the object search server 2 also returns a plurality of reference image IDs to the terminal 1 together with the similarity determination result.
Next, for S5 (geometric information transmission unit of the terminal), the terminal 1 also transmits the previously received reference image ID together with the geometric information.
And about S6 (geometric verification part of an object search server), the object search server 2 should just perform the process of geometric verification among several reference image ID received from the terminal 1. FIG.

  As described above in detail, according to the system and method of the present invention, when searching for an object shown in a captured image, local features transferred via a network can be transmitted without reducing the object search accuracy. Data size can be reduced.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Terminal 11 Local feature extraction part 12 Quantization part 13 Geometric information storage part 14 Question information transmission part 15 Geometric information selection part 16 Geometric information transmission part 2 Object search server 21 Local feature extraction part 22 Quantization part 23 Geometric information storage part 24 Transposed index storage unit 25 Similarity calculation unit 26 Similarity determination transmission unit 27 Geometric verification unit 28 Search result transmission unit

Claims (8)

An object search method for a system having a terminal and an object search server,
The object search server stores in advance a reference image showing an object,
A first step in which the terminal transmits question information including a plurality of feature vectors in a question image in which an object is captured to the object search server;
A second step in which the object search server calculates a similarity between a plurality of feature vectors of the question information and each reference image;
A third step in which the object search server returns a similarity determination result between the question information and a plurality of reference images to the terminal;
A fourth step in which the terminal selects geometric information including coordinate information of the question image according to the received similarity determination result;
A fifth step in which the terminal transmits the selected geometric information to the object search server;
A sixth step in which the object search server searches for one or more reference images similar to the coordinate information of the geometric information at a predetermined threshold or more among the plurality of reference images searched in the second step;
And a seventh step of returning the reference image ID of each reference image searched in the fifth step to the terminal. For the first step, the terminal
Extracting a local feature including a feature vector and coordinate information from a query image showing an object;
Quantizing the feature vector into a quantization vector;
Transmitting question information including a plurality of quantization vectors to the object search server,
For the second step, the object search server
The object search method according to claim 1, wherein the similarity between the plurality of quantization vectors of the question information and each reference image is calculated. The object search method according to claim 2, wherein the quantization vector included in the question information is a Visual Word identifier. The object search method according to claim 3, wherein the question information further includes a direction and / or a scale in association with a Visual Word identifier. For the third step, the object search server returns one of the following three commands as the similarity determination result to the terminal:
yes1: There is one or more reference images whose similarity is greater than or equal to the threshold TH1
yes2: There is one or more reference images whose similarity is less than the threshold TH1 and greater than or equal to the threshold TH2.
no: all similarities are less than TH2 threshold TH1: index of similarity with very high possibility of successful search threshold TH2: index of similarity with high possibility of successful search For the fourth step, the terminal However, according to the similarity determination result, the geometric information to be transmitted to the object search server is controlled as follows.
yes1: Geometric information includes only coordinate information
yes2: Geometric information includes coordinate information and feature vectors
The object search method according to any one of claims 2 to 4, wherein the process is terminated as a search failure. For the third step, the object search server returns, as the similarity determination result, any one of four or more commands to the terminal according to three or more threshold values,
The fourth step is characterized in that, as the threshold based on the similarity determination result is lower, the terminal includes a longer feature vector (only upper bits of the feature vector-> total bits) in the geometric information. The object search method described in 1. For the sixth step, the object search server
A feature point correspondence set that satisfies a predetermined geometric condition is calculated between the coordinate information of the geometric information and the coordinate information of each reference image,
The object search method according to any one of claims 1 to 6, wherein a reference image having a maximum number of feature point correspondence groups or a predetermined number or more is searched. A system having a terminal and an object search server,
The terminal
Question information transmitting means for transmitting question information including a plurality of feature vectors in a question image in which an object is shown to the object search server;
Geometric information selection means for selecting geometric information including coordinate information of the question image according to the received similarity determination result;
Geometric information transmission means for transmitting the selected geometric information to the object search server;
The object search server stores in advance a reference image showing an object,
Similarity calculating means for calculating the similarity between the plurality of feature vectors of the question information and each reference image;
Similarity determination transmission means for returning a similarity determination result between the question information and the plurality of reference images to the terminal;
Geometric verification means for searching for one or more reference images that are similar to the coordinate information of the geometric information at a predetermined threshold value or more among a plurality of reference images determined to be similar based on the similarity determination result;
And a search result transmitting means for returning a reference image ID of each reference image searched by the geometric verification means to the terminal.

Images