JP2012160047A - Corresponding reference image retrieval device and method thereof, content superimposing apparatus, system, method, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a corresponding reference image retrieval device capable of effectively retrieving a reference image corresponding to an input image in an apparatus having limited resources.SOLUTION: A corresponding reference image retrieval device 10 for retrieving a reference image corresponding to an input image includes: a feature amount detection unit 12 which extracts a feature point from the input image and detects a feature amount of the feature point; a binary conversion unit 13 which converts the feature amount detected in the feature amount detection unit 12 to a binary code; a feature point database 14 which stores the feature amounts for each of the feature points of a plurality of reference images in a format of binary code; a matching unit 15 which detects the reference image corresponding to the input image from the plurality of reference images by comparing the feature amount of the binary code for the input image converted in the binary conversion unit 13 with the feature amount of the binary code for the plurality of reference images stored in the feature point database 14.

Description

  The present invention relates to a corresponding reference image search apparatus and method for searching for a reference image corresponding to an input image, and a content superimposing apparatus, system, and method for superimposing corresponding content on an input image using them, and in particular, BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a corresponding reference image search apparatus and method for searching a reference image corresponding to an input image using feature points of the image, and a content superimposing apparatus, system, and method for superimposing corresponding content on the input image using them. Is.

  In recent years, with the widespread use of mobile terminals with cameras, AR (Augmented Reality) that captures a target object with a camera as an input image and displays content corresponding to the target object (for example, explanation of the target object) superimposed on the input image. Technology has been proposed.

  As conventional techniques for realizing the AR, there are one that does not perform image processing and one that performs image processing in order to specify content corresponding to a target object. As a prior art that does not perform image processing, there is one that uses a GPS receiver and an electronic compass in combination. In this prior art, the position of the target object is stored in advance in a database, the position of the mobile terminal with camera is detected by a GPS receiver, the orientation of the mobile terminal with camera is detected by an electronic compass, and the position and orientation are determined. Based on this, by estimating the target object within the angle of view of the camera and its position, the content corresponding to the target object and its overlapping location are specified.

  The conventional technology that does not perform image processing to identify the content corresponding to the target object can be realized relatively easily if the camera-equipped mobile terminal is equipped with a GPS receiver and an electronic compass. Since it depends on the accuracy of the GPS receiver and the electronic compass, it is difficult to accurately superimpose the content on the target object.

  On the other hand, the conventional technology that performs image processing to identify the content corresponding to the target object shown in the input image identifies the content corresponding to the target object by recognizing the target object shown in the input image. . As a conventional technique for performing image processing, one using an index and one using a corresponding point have been proposed.

  The technique using an index uses, for example, a two-dimensional code as an index, stores content and a corresponding two-dimensional code in advance in a database, and assigns a two-dimensional code to the target object or its vicinity. The content corresponding to the two-dimensional code photographed from the database is searched by photographing the target object including the two-dimensional code with the camera-equipped mobile terminal.

  According to this prior art, it is possible to estimate the attitude of the camera-equipped mobile terminal with respect to the plane to which the two-dimensional code is assigned at high speed and with high definition. In addition, by adopting a two-dimensional code, a large number of code patterns can be easily created, so that various objects can be recognized. However, since it is necessary to add a two-dimensional code to the target object or its vicinity, it is impractical to superimpose content on, for example, an outdoor landmark or a large signboard. Moreover, the influence with respect to the design of a target object becomes large by giving a two-dimensional code to a target object or its vicinity.

  The technique using the corresponding points searches for the reference image corresponding to the input image by obtaining the corresponding point between the input image and the reference image in the database, and superimposes the content corresponding to the reference image on the input image. To do. By using corresponding points, a markerless AR that does not require an index such as a two-dimensional code can be realized, so that the application range is widened and the design of the target object is not affected.

  In the technique of searching for a corresponding reference image using corresponding points, a plurality of feature points of the input image are extracted, the feature amount of each feature point (local feature amount) is compared with the feature amount of the feature point of the reference image, Corresponding feature points (corresponding points) are searched, and a reference image including many corresponding feature points is set as a reference image corresponding to the input image. As such a technique, there are known image feature point expression methods such as SIFT (Scale-invariant Feature Transform) and SURF (Speeded Up Robust Features) obtained by increasing the speed of SIFT (see Non-Patent Document 1). These image feature point expression methods have an advantage that recognition ability is high and various objects can be recognized. In addition, since a list of corresponding points between the input image and the reference image is obtained, it is possible to easily execute calculations necessary for superimposing content on the input image.

David G. Lowe, "Object recognition from local scale-invariant features," International Conference on Computer Vision, Corfu, Greece (September 1999), pp. 1150-1157

  However, the above-described conventional technique for searching for a reference image corresponding to an input image by comparing feature amounts of feature points has the following problems. First, in the above-described conventional technology, the size of the feature amount is large. Since feature quantities are expressed as vectors of hundreds of dimensions expressed as single-precision real numbers, if hundreds to thousands of feature points are extracted from one image, the data amount of the feature quantity is from tens of kilobytes to It can be several megabytes. Therefore, it is difficult to maintain a database in the main storage area in an apparatus with limited hardware resources such as a portable terminal.

  Further, in the above-described conventional technology, it is necessary to calculate the L2 norm between the feature amounts expressed by vectors. The calculation of the distance between these high-dimensional vectors is extremely expensive. Therefore, it is difficult to realize with an apparatus with limited computing resources such as a portable terminal.

  The present invention has been made in view of the above problems, and a corresponding reference image retrieval apparatus and method capable of retrieving a reference image corresponding to an input image effectively even in an apparatus with limited resources, It is another object of the present invention to provide a content superimposing apparatus, system, and method for superimposing corresponding content on an input image using them.

  In order to solve the above-described conventional problems, a corresponding reference image search device according to the present invention is a corresponding reference image search device that searches for a reference image corresponding to an input image, and extracts feature points from the input image to extract the feature points. A feature amount detection unit that detects a feature amount of a point; a binary conversion unit that converts the feature amount detected by the feature amount detection unit into a binary code; and a feature amount of each feature point of a plurality of reference images Feature point database stored in code format, binary code feature amount of the input image converted by the binary conversion unit, and binary code feature amount of the plurality of reference images stored in the feature point database And a matching unit that detects a reference image corresponding to the input image from the plurality of reference images.

  With this configuration, the feature values of a plurality of reference images are stored as binary data, and the feature values of binary codes are compared even when searching for corresponding reference images. A reference image corresponding to the image can be searched.

  In the corresponding reference image search device, the binary conversion unit may convert the feature quantity detected by the feature quantity detection unit into a binary code using a transformation matrix.

  With this configuration, the calculation cost of conversion in the binary conversion unit can be reduced.

  In the corresponding reference image search device, the conversion matrix may be a sparse matrix.

  With this configuration, the calculation cost of conversion in the binary conversion unit can be further reduced.

  In the corresponding reference image search device, the binary conversion unit may be able to change the size of the binary code by changing the size of the conversion matrix.

  With this configuration, it is possible to make adjustments such as reducing the calculation cost in the matching unit by reducing the size of the binary code and improving the search accuracy in the matching unit by increasing the size of the binary code.

  In the corresponding reference image search device, the matching unit may detect a plurality of reference images when there are a plurality of reference images corresponding to the input image.

  With this configuration, when a plurality of target objects are reflected in the input image, a plurality of reference images are detected.

  In addition, the corresponding reference image search device may further include an environment measurement unit that measures an execution environment of the corresponding reference image search device, and the binary conversion unit converts the conversion according to a measurement result by the environment measurement unit. The size of the binary code may be changed by changing the size of the matrix.

  With this configuration, the binary conversion unit changes the feature amount detected by the feature amount detection unit into a binary code having a size according to the execution environment (for example, the capacity of the storage unit, the capacity of the calculation unit, and the calculation processing capacity). it can.

  A content superimposing apparatus of the present invention is a content superimposing apparatus that includes the corresponding reference image search device described above and superimposes the corresponding content on the input image, and the correspondence between the content and the reference image and the content A content database that stores the relationship, a content extraction unit that extracts content corresponding to the reference image detected by the matching unit from the content database, and a content that is extracted by the content extraction unit is superimposed on the input image And a superimposing unit.

  With this configuration, it is possible to effectively superimpose the corresponding content on the input image even in an apparatus with limited resources.

  In the content superimposing apparatus, the feature amount detection unit may extract feature points including position information in the input image, and the feature point database may store feature points of a plurality of reference images. Information on the position of each feature point may be stored together with the feature amount, the content database may further store a superimposed position of the content, and the content superimposing device may further detect the feature amount. The content extracted by the content extraction unit based on the relationship between the position of the feature point extracted by the unit and the position of the feature point stored in the feature point database is stored in the content database. The content conversion part which converts a superimposition position may be provided, and the said superimposition part is the superimposition converted by the said content conversion part in the said input image. It may be superimposed the contents extracted by the contents extraction unit to location.

  With this configuration, even if the target position on which the content is superimposed is different between the input image and the reference image, the content can be superimposed at an appropriate position in the input image.

  The content superimposing system of the present invention is a content superimposing system comprising a content superimposing device and an external search server that can communicate with the content superimposing device. The content superimposing apparatus extracts a feature point from an input image and detects a feature amount of the feature point, and binary conversion that converts the feature amount detected by the feature amount detection unit into a binary code And a content superimposing device side communication unit that transmits the feature value of the binary code of the input image converted by the binary converting unit to the external search server, and the external search server is the content superimposing device side An external search server side communication unit that receives a binary code feature amount of the input image transmitted from the communication unit, and an external search server side that stores the feature amount of each feature point of a plurality of reference images in the form of a binary code Feature point database, feature quantity of binary code of the input image received by the external search server side communication unit, and feature point data of the external search server An external search server-side matching unit that detects a reference image corresponding to the input image from the plurality of reference images by comparing the feature values of binary codes of the plurality of reference images stored in a database; It has the composition provided with.

  With this configuration, the content superimposing device transmits a binary code as a feature quantity of the input image to the external search server. Therefore, compared with a case where a feature quantity having a large amount of data such as a single precision real number is transmitted as it is, transmission is performed. Data volume can be reduced. Also, the external search server can reduce the calculation cost and the required database capacity.

  In the content superimposing system, the feature amount detection unit may extract feature points including position information in the input image, and the feature point database may store feature points of a plurality of reference images. Information on the position of each feature point may be stored together with the feature amount. The content superimposition system detects the correspondence between the reference image stored in the feature point database and the content, the content database storing the content superimposition position, and the external search server side matching unit. Extracting content corresponding to a reference image from the content database, based on the relationship between the position of the feature point extracted by the feature amount detection unit and the position of the feature point stored in the feature point database, A content conversion unit that converts the superposition position stored in the content database of the content extracted from the content database, and the superposition position converted by the content conversion unit in the input image by the content conversion unit And a superimposing unit that superimposes the extracted content. There may be.

  With this configuration, it is possible to superimpose corresponding content on the input image, and even if the target position on which the content is superimposed differs between the input image and the reference image, the content is placed at an appropriate position in the input image. Can be superimposed.

  Further, in the content superimposing system, the content superimposing device includes a feature superimposing device-side feature point database in which feature amounts of feature points of a plurality of reference images are stored in a binary code format, and the binary converting unit. By comparing the converted binary image feature quantity of the input image with the binary code feature quantity stored in the content superimposing apparatus side feature point database, the input image is selected from the plurality of reference images. A content superimposing device side matching unit that detects a reference image corresponding to the external search server side communication unit, wherein the external search server side communication unit includes the external search server among the feature quantities stored in the external search server side feature point database. The reference image detected by the side matching unit and the feature amount of the reference image related thereto are used as the content weight. The content superimposing device side communication unit may receive the feature amount of the binary code transmitted from the external search server side communication unit, and the content superimposing device side feature point database The feature amount of the binary code received by the content superimposing apparatus side communication unit may be the feature amount of each feature point of the plurality of reference images.

  With this configuration, the content superimposing apparatus can download only necessary data from the external search server even if it does not store all of the feature quantities of a large amount of reference images. Database capacity can be reduced.

  Further, the corresponding reference image search method of the present invention corresponds to an input image in a corresponding reference image search device including a feature point database in which feature quantities of feature points of a plurality of reference images are stored in a binary code format. A corresponding reference image retrieval method for retrieving a reference image, the feature point extracting step for extracting a feature point from an input image, and the feature amount detection for detecting the feature amount of the feature point extracted in the feature point extracting step A binary code conversion step for converting the feature amount detected in the feature amount detection step into a binary code, a feature amount of the binary code of the input image converted in the binary code conversion step, and the feature The plurality of reference images are compared by comparing the binary code feature amount of each of the plurality of reference images stored in the point database. And a matching step of detecting a reference image corresponding to the input image from within.

  Also with this configuration, the feature amounts of a plurality of reference images are stored as binary data, and the feature amounts of the binary codes are compared even in the search for the corresponding reference images. Therefore, even in an apparatus with limited resources, A reference image corresponding to the input image can be searched.

  Also, the content superimposing method of the present invention includes a corresponding reference image search device provided with a feature point database in which feature amounts of feature points of a plurality of reference images are stored in a binary code format, and stored in the feature point database. A content superimposing method for superimposing corresponding content on an input image in a content superimposing apparatus provided with a content database storing a correspondence relationship between the reference image and the content. A corresponding reference image search step for detecting a reference image corresponding to the input image, a content extraction step for extracting content corresponding to the reference image detected in the corresponding reference image search step from the content database, and the content The content extracted by the extraction unit is superimposed on the input image. And a tatami step.

  With this configuration, it is possible to effectively superimpose the corresponding content on the input image even in an apparatus with limited resources.

  The computer program of the present invention is a computer program for causing a computer to execute the above-described corresponding reference image search method.

  A computer program according to another aspect of the present invention is a computer program for causing a computer to execute the above-described content superimposing method.

  According to the present invention, the feature values of a plurality of reference images are stored as binary data, and the feature values of the binary codes are compared even when searching for the corresponding reference images. Therefore, it is effective even in an apparatus with limited resources. The reference image corresponding to the input image can be searched.

The block diagram which shows the structure of the content superimposition apparatus in embodiment of this invention. The figure which shows the example of the input image in embodiment of this invention The figure which shows the feature point extracted from the input image in embodiment of this invention The figure which shows the feature-value detected from the input image in embodiment of this invention The figure which shows the feature-value converted into the binary code in embodiment of this invention The figure which shows the data preserve | saved in the feature point database in embodiment of this invention The figure which shows the corresponding point pair in embodiment of this invention The figure which shows the data preserve | saved in the content database in embodiment of this invention The block diagram which shows the structure of the content superimposition apparatus in the modification 2 of embodiment of this invention. The block diagram which shows the structure of the content superimposition apparatus in the modification 3 of embodiment of this invention. The block diagram which shows the structure of the content superimposition apparatus in the modification 4 of embodiment of this invention. The figure which shows the structure of the data preserve | saved in the feature point database of the external search server in the modification 4 of embodiment of this invention.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration of the content superimposing apparatus according to the present embodiment. The content superimposing device 110 includes a corresponding reference image search device 10 for searching for a reference image corresponding to an input image. The content superimposing device 110 uses the reference image searched by the corresponding reference image search device 10 to As a configuration for superimposing related content, a correspondence calculation unit 21, a content conversion unit 22, a content database 23, and a superposition unit 24 are provided.

  The corresponding reference image search apparatus 10 includes an image acquisition unit 11, a feature amount detection unit 12, a binary conversion unit 13, a feature point database 14, and a matching unit 15. The image acquisition part 11 produces | generates an image by image | photographing with the camera as an imaging device, and acquires this as an input image. The image acquisition unit 11 may input an image generated outside through communication or a recording medium. FIG. 2 is a diagram illustrating an example of an input image. In the following description, processing in each unit will be described using the input image of FIG. The input image acquired by the image acquisition unit 11 is output to the feature amount detection unit 12.

  The feature amount detection unit 12 extracts feature points from the input image and detects feature amounts of the extracted feature points. FIG. 3 is a diagram illustrating feature points extracted from the input image. As shown in FIG. 3, generally, a plurality of feature points are detected from the input image.

  FIG. 4 is a diagram illustrating the feature amounts detected by the feature amount detection unit 12. In the present embodiment, a local feature amount is used as the feature amount. Specifically, a SIFT feature value is used as the feature value. In practice, other local feature quantities such as SURF feature quantities may be used. As shown in FIG. 4, in the feature quantity detection unit 12, the local feature quantity is obtained as a single-precision real vector. The feature quantity detection unit 12 outputs the position information of each feature point and the local feature quantity detected for each feature point to the binary conversion unit 13.

The binary conversion unit 13 converts the feature amounts of all feature points extracted from the input image into binary codes. FIG. 5 is a diagram illustrating the feature amount converted into the binary code. Assuming that the feature quantity detected by the feature quantity detection unit 12 is a 128-dimensional vector vεR ¹²⁸ , the binary conversion unit 13 converts this feature quantity into a binary code by the following equation (1).

また、ベクトルｗ_kは、１２８次元における半径１の超球上の点から、一様分布に従ってランダムサンプリングをして得られるベクトルである。ｗ_k（ｋ＝１，……ｄ）は、１２８行ｄ列の行列として表現できる。このベクトルｗ_kからなる行列を「ｗ」と表記し、変換行列という。 In equation (1), d is the size (that is, the number of bits) of the binary code after conversion, and the sgn function is given by the following equation (2).

The vector w _k is a vector obtained by random sampling according to a uniform distribution from a point on a hypersphere having a radius of 1 in 128 dimensions. w _k (k = 1,... d) can be expressed as a matrix of 128 rows and d columns. A matrix composed of the vector w _k is expressed as “w” and is referred to as a transformation matrix.

Before executing the above conversion, the binary conversion unit 13 samples feature amounts from a large amount of images in advance, subtracts the average or median m obtained therefrom from vεR ¹²⁸ , and VεR ¹²⁸ is normalized so that the L2 norm is 1. In the present embodiment, m is generated using a large amount of feature amounts generated when data is stored in the feature point database 14. In this embodiment, the bit length of the binary code is 128 bits, that is, d = 128.

  The feature point database 14 stores feature amounts of feature points of a plurality of reference images in a binary code format. This reference image is a recognition target image in the AR. FIG. 6 is a diagram showing data stored in the feature point database 14. As shown in FIG. 6, in the feature point database 14, for each feature point, the image identification number of the reference image to which the feature point belongs, the feature amount of the feature point expressed in binary code, and the feature point A record consisting of a position in the image is stored.

  These records stored in the feature point database 14 are prepared using the image acquisition unit 11, the feature amount detection unit 12, and the binary conversion unit 13 described above. That is, the image acquisition unit 11 acquires a reference image by photographing or reading data from a network or a recording medium, and the feature amount detection unit 12 extracts a feature point from the reference image, Detect feature values. The feature amount detection unit 12 assigns an image identification number to the reference image, and outputs the position information of each feature point in the reference image and the detected feature amount to the binary conversion unit 13 together with the image identification number. The binary conversion unit 13 performs binary conversion on the feature value to generate a binary code.

  As shown in FIG. 6, a plurality of feature points are extracted from one reference image. In the example of FIG. 6, only a few feature point records are shown for one reference image, but in practice hundreds to thousands of feature point records may be stored for one reference image. In the example of FIG. 6, only two reference images are shown, but the feature point database 14 may store feature point records for thousands or more reference images. Furthermore, in the example of FIG. 6, for convenience of space, the binary code shows only the first 11 digits. However, as described above, in this embodiment, the binary conversion unit 13 generates a 128-bit binary code. The feature point database 14 also stores a 128-bit binary code.

  The matching unit 15 compares the binary code indicating the feature amount of the feature point extracted from the input image with each of the binary code indicating the feature amount of the feature point of the reference image stored in the feature point database 14. Search for the nearest binary code. In the present embodiment, the matching unit 15 employs a Hamming distance for evaluation of the perspective of the binary code.

Note that the Hamming distance of the binary code coincides with the cosine distance in the original input vector space in the limit. That is, when any two vectors v ₁ and v ₂ are converted by h _k (v), the probability that each bit becomes a different value is proportional to the angle formed by the vectors v ₁ and v _2, and Equation (3) holds.

  The value of the probability on the left side of Equation (3) can be regarded as the Hamming distance of the binary code. Therefore, if a sufficiently long bit string is obtained, the cosine distance of the vector in the original space matches the Hamming distance of the binary code. Therefore, the matching unit 15 may evaluate the distance between the binary codes using Expression (3). The matching unit 15 may search for the nearest binary code using an existing technique such as LSH (Locality Sensitive Hashing) according to the purpose of speeding up.

  The matching unit 15 compares the binary code of all the records stored in the feature point database 14 with the binary code of the feature amount of the feature point extracted from the input image, and determines the feature point having the closest binary code. Vote against it. As a result of voting on all the feature points extracted from the input image, the matching unit 15 obtains the reference image that has acquired the most votes as a reference image corresponding to the input image (hereinafter referred to as “corresponding reference image”). .)

  The matching unit 15 may set a lower limit of the number of acquired votes for determining that it is a corresponding reference image. In this case, if the number of acquired votes of the reference image that has acquired the most votes is less than this lower limit, it is determined that no corresponding reference image exists in the feature point database 14. Further, the matching unit 15 may set a threshold value for the number of acquired votes for determining that it is a corresponding reference image, and may set all reference images that have acquired votes equal to or greater than the threshold value as corresponding reference images.

  The matching unit 15, together with the image identification number of the corresponding reference image, information on the position of the feature point voted in the corresponding reference image (this feature point is referred to as “corresponding point of the corresponding reference image”), the feature point of the input image Information of the position of the feature point voted for the corresponding reference image (this feature point is referred to as “corresponding point of the input image”) is output to the correspondence calculating unit 21. At this time, the matching unit 15 outputs the corresponding points of the corresponding reference image and the corresponding points of the input image for which voting has been performed as a pair. FIG. 7 is a diagram showing a pair of corresponding points.

このホモグラフィ行列Ａを用いると、対応参照画像上の任意の点を入力画像上の点に写像させることができる。対応関係算出部２１は、対応参照画像の画像識別番号とホモグラフィ行列Ａとをコンテンツ変換部２２に出力する。 The correspondence calculation unit 21 calculates a homography matrix that maps an arbitrary point (coordinate) on the corresponding reference image to a point (coordinate) on the input image based on the plurality of corresponding point pairs input from the matching unit 15. To do. Specifically, the correspondence calculation unit 21 estimates the homography matrix A satisfying the following equation (4) by the RANSAC (Random Sample Consensus) method using the corresponding point pair input from the matching unit 15.

If this homography matrix A is used, an arbitrary point on the corresponding reference image can be mapped to a point on the input image. The correspondence calculation unit 21 outputs the image identification number of the corresponding reference image and the homography matrix A to the content conversion unit 22.

  The content database 23 stores content to be superimposed on the input image. FIG. 8 is a diagram showing data stored in the content database 23. As shown in FIG. 8, the content database 23 stores, for each content, a record including an image identification number of a reference image corresponding to the content, content data, a shape, a size, and a superimposition location of the content.

  The content data may be various data superimposed on the input image, including text data, image data, and moving image data. In the example of FIG. 8, the reference image with the image identification number 1 includes, as content data, a description (text data) of the target object shown in the reference image “This photo is ...”, “http: /// A reference URL (text data) “www.abcdefg.com” and image data of a reference image are prepared.

  The content conversion unit 22 extracts content corresponding to the image identification number of the corresponding reference image input from the correspondence calculation unit 21 from the content database 23. At this time, the content conversion unit 22 functions as a content extraction unit of the present invention. The content conversion unit 22 converts the superimposed position of the extracted content using the homography matrix A input from the correspondence calculation unit 22 and outputs the converted content to the superposition unit 24 together with the content data.

  The superimposing unit 24 acquires the input image from the image acquiring unit 11, superimposes the content data obtained from the content converting unit 22 on the input image, and outputs it. At this time, the superimposing unit 24 superimposes the content data on the superposed position after conversion output from the content converting unit 22 in the input image.

  As described above, according to the corresponding reference image retrieval apparatus 10 of the present embodiment, matching between the input image and the reference image (retrieval of the corresponding reference image) is performed using the feature amount of the binary code. The burden of processing can be reduced. In addition, since the database storing the feature quantities of the feature points of the reference image also stores the feature quantities in the form of binary code, the capacity required for the database can be small. Therefore, the corresponding reference image search apparatus 10 can be realized in a limited resource apparatus, and the content superimposing apparatus 110 including the corresponding reference image search apparatus 10 can also be realized in a limited resource apparatus.

  The present invention is not limited to the above-described embodiment, and various modifications can be made. Hereinafter, modified examples will be described.

(Modification 1)
The binary conversion unit 13 of the above embodiment converts the feature quantity into a binary code using the above equation (1). At this time, the vector w _k can be made sparse. Further, if the vector w _k is not sampled from a uniform distribution but is sampled as in the following equation (5) to make w a sparse matrix, equation (1) is approximately established.

  Such conversion is called a very sparse random mapping (Very Sparse Random Projection). Since w at this time is a sparse matrix and non-zero elements are composed of only -1 or +1, multiplication is not necessary in the calculation of the matrix w, and the number of additions / subtractions can be very small. . Therefore, the calculation cost in the binary conversion unit 13 can be greatly reduced.

(Modification 2)
FIG. 9 is a block diagram illustrating a configuration of a content superimposing apparatus according to the second modification. In the content superimposing device 120 of Modification 2, the environment measuring unit 16 is added to the corresponding reference image search device 20 as compared to the content superimposing device 110 of the above-described embodiment. Further, the binary conversion unit 13 of the corresponding reference image search device 20 of the content superimposing device 120 can change the size of the binary code generated by the conversion.

  The corresponding reference image search device or content superimposing device of the present invention is mounted on various devices such as a mobile phone terminal and a notebook computer. Therefore, the processing speed of the CPU, which is a resource of the device, and the capacity of the database (main storage device) are different for each execution environment. On the other hand, as the size (number of bits) of the binary code generated by the binary conversion unit 13 is smaller, the calculation cost can be reduced and the necessary database capacity can be reduced. However, on the other hand, the binary code representing the feature amount is used. If the size of is small, the accuracy of matching (search for corresponding reference image) becomes low.

  Therefore, the environment measurement unit 16 measures the processing speed of the CPU and the capacity of the database of the device on which the corresponding reference image search device 20 or the content superimposing device 120 is mounted, and generates the binary conversion unit 13 according to the measurement result. Determine the size of the binary code to be executed. Then, based on the size determined by the environment measurement unit 16, the binary conversion unit 13 converts the single-precision real number feature quantity into a binary code.

  Further, the binary conversion unit 13 adjusts the size of the matrix w used when generating the binary code as the feature amount of the reference image to be stored in the feature point database 14, thereby reducing the size of the generated binary code. adjust.

  For example, if the single-precision real feature vector detected by the feature detector 12 is D-dimensional and the binary code stored in the feature point database 14 is 128 bits, the binary converter 13 In order to obtain the feature amount of the feature point of the reference image to be stored in the point database, the single-precision real number feature amount is converted into a binary code using the matrix w of 128 rows and D columns.

  In this case, when the size of the binary code determined by the environment measurement unit 16 is 64 bits, the binary conversion unit 13 cuts out the first row to the 64th row from the matrix w of 128 rows and D columns. , A matrix w ′ having a size of 64 rows and D columns is generated, and the feature amount of the feature point of the input image is converted into a binary code using the matrix W ′. In this case, the matching unit 15 compares the binary code (64 bits) that is the feature amount of the feature point of the input image with the upper 64 bits of the binary code (128 bits) stored in the feature point database 14. Then, matching is performed by determining the feature points to vote.

  According to the corresponding reference image search device 20 and the content superimposing device 120 of the second modification, the environment measurement unit 16 causes the processing speed of the CPU of the device on which the corresponding reference image search device 20 or the content superimposing device 120 is mounted, and the database Since the size of the binary code generated by the binary conversion unit 13 is determined according to the execution environment such as the capacity, a binary code suitable for the execution environment can be generated.

(Modification 3)
The corresponding reference image search device or the content superimposing device may be mounted on a device that can always communicate with an external network, such as a mobile phone terminal. When the corresponding reference image search device or the content superimposing device can always communicate with an external network, the corresponding reference image can be searched by an external device.

  FIG. 10 is a block diagram illustrating a configuration of a content superimposing system according to the third modification. The content superimposing system 101 includes a content superimposing device 130 and an external search server 230. The content superimposing apparatus 130 and the external search server 230 are provided with communication units 31 and 41, respectively, and communicate with each other.

  The configuration of the content superimposing apparatus 130 is the same as that of the content superimposing apparatus 110 of the above embodiment, except that it includes the communication unit 31 and the size of the binary code generated by the binary conversion unit 13 can be changed. It is. In addition to the communication unit 41, the external search server 230 includes a matching unit 42, a feature point database 43, and a content database 44.

  The content superimposing apparatus 130 can retrieve the corresponding reference image and superimpose the content stored in the content database 23 on the input image in the same manner as in the above embodiment. The content superimposing apparatus 130 can further transmit a binary code representing the feature amount of the feature point extracted from the input image to the external search server 230 via the communication unit 31.

  The external search server 230 receives the feature value of the binary code by the communication unit 41. The matching unit 42 searches for the corresponding reference image by voting on the feature points stored in the feature point database 43 based on the binary code. The communication unit 41 transmits the image identification number of the corresponding reference image and the information on the corresponding point pair to the content superimposing apparatus 130. The communication unit 41 also extracts a content record (including content data and content superposition position information) corresponding to the corresponding reference image from the content database 44 and transmits the content record to the content superposition device 130.

  The communication unit 31 of the content superimposing apparatus 130 receives the information on the corresponding point pair and the content record corresponding to the corresponding reference image. The correspondence calculation unit 21 calculates a homography matrix based on the corresponding point pair received by the communication unit 31. The content changing unit 22 converts the superimposed position of the content received by the communication unit 31 using the homography matrix calculated by the correspondence calculating unit 21. Similar to the above embodiment, the superimposing unit 24 superimposes the content converted from the superimposition position output from the content converting unit 22 on the input image obtained from the image acquiring unit 11.

  In Modification 3, when the corresponding reference image is searched by the external search server 230 as described above, the binary conversion unit 13 changes the size of the binary code according to the calculation capability of the external search server 230. To do. When the external search server 230 has a higher function than the content superimposing apparatus 130, the binary conversion unit 13 increases the size of the binary code (increases the number of bits).

  According to the third modification example, the corresponding reference image is not searched only in an environment where the calculation resources are limited, such as a mobile phone terminal, but the search can be performed by an external search server with a better calculation environment. . Moreover, since the binary conversion unit 13 can adjust the size of the binary code to an appropriate size, the small-scale real-time matching processing is performed in the content superimposing apparatus 130 that is a mobile phone terminal as in the above embodiment. When searching for a corresponding reference image and performing a large-scale matching process, an external search server having a higher processing capacity than a mobile phone terminal should search for the corresponding reference image using a binary code having a larger size. Can do.

  When the content superimposing device 130 does not superimpose content alone, the content superimposing device 130 may not include the feature point database 14, the matching unit 15, or the content database 23, and the correspondence relationship. The external search server 230 may include the calculation unit 21, the correspondence relationship calculation unit 21 and the content conversion unit 22, or the correspondence relationship calculation unit 21, the content conversion unit 22, and the superposition unit 24. Conversely, the external search server 230 may not include the content database 44 and only the result of matching by the matching unit 42 may be transmitted to the content superimposing device 130.

(Modification 4)
When the content superimposing apparatus includes a communication unit and communicates with the external search server, the following content superimposing system can also be realized. FIG. 11 is a block diagram illustrating a configuration of a content superimposition system according to the fourth modification. The content superimposing system 102 includes a content superimposing device 140 and an external search server 240. This content superimposing system 102 is preferably used when the target of superimposing content is an image of a book page. Below, the content superimposition system 102 is demonstrated to an example when the object on which a content is superimposed is the image of the page of a book.

  For example, if there are 5000 books to store reference images, and the average number of pages of each book is 300 pages, the feature point database includes 1.5 million page feature points (a plurality of features for each page). It is necessary to save the record. However, it is impossible to store all of such a large amount of data in a small or portable device such as a mobile phone terminal due to the capacity limitation of the storage device. Even if such a large amount of data can be stored in the mobile phone terminal, there are too many search targets, and the calculation cost of matching becomes enormous. Furthermore, since there are many search objects, there is a problem that the search accuracy deteriorates unless the binary code is lengthened.

  Therefore, in the content superimposing system 102, the records of all the feature points for all the pages of all the books can realize a large-capacity database relatively easily, and the external search server 240 has relatively few physical restrictions. It is stored in the feature point database 43.

  FIG. 12 is a diagram showing a configuration of data stored in the feature point database 43. As shown in FIG. 12, the feature point database 43 stores a record including a book number, a page number, a binary code (128 bits and 64 bits), and a feature point position (coordinates) for each feature point. Yes. Similar to the above embodiment, records of a plurality of feature points are stored for one reference image (one-page image).

  In the content superimposing apparatus 140, when the image acquisition unit 11 acquires an input image, the feature amount detection unit 12 detects the feature amount of the feature point, and the binary conversion unit 13 converts the feature amount into a binary code. Then, the communication unit 31 transmits the binary code to the external search server 240.

  The external search server 240 receives the binary code from the content superimposing device 140 at the communication unit 41. The matching unit 42 performs voting using the binary code received by the communication unit 41 (the number of feature points extracted from the input image), and detects the reference image that has acquired the most votes as a corresponding reference image. To do. The corresponding reference image is a one-page image of one book out of 5000 books.

  Among the records stored in the feature point database 43, the communication unit 41 records the book number to which the corresponding reference image detected by the matching unit 42 belongs (that is, other reference related to the corresponding reference image and the corresponding reference image). Image record) is transmitted to the content superimposing apparatus 140. At this time, the communication unit 41 transmits at least a page number, a binary code (either one of 128 bits and 64 bits), and feature point position information in the book number record. Data downloaded to the content superimposing device 140 from the external search server 240 in this way is referred to as feature point database update data.

  Whether the 128-bit or 64-bit binary code is transmitted depends on the transmission destination, that is, the calculation capability of the content superimposing apparatus 140 and the capacity of the database. Whether the 128-bit or 64-bit binary code is to be transmitted may be determined by the external search server 240 recognizing the calculation capability of the content superimposing device 140 that is the transmission destination or the capacity of the database. The size of the binary code may be specified from the 140 side.

  As described above, in order to specify a book in the external search server 240, the content superimposing apparatus 140 first transmits the feature amount of the feature point of the input image to the external search server 240. This feature amount (binary data) ) And the size of the binary code transmitted from the external search server 240 to the content superimposing device 140 are not necessarily the same. In particular, the former may be large and the latter may be small.

  The communication unit 31 of the content superimposing apparatus 140 that has received the feature point database update data, which is a part of the records in the feature point database 43 of the external search server 240, stores this in the feature point database 14 or already has the feature point database. If any data is stored in the feature point database 14, the feature point database 14 is updated with the feature point database update data. For subsequent input images, matching (search for corresponding reference images) is performed using the feature point database update data stored in the feature point database 14. At this time, the binary conversion unit 13 converts the single-precision real number feature quantity detected by the feature quantity detection unit 12 into a binary code so as to have the same size as the binary code size in the feature point database update data.

  According to the content superimposing system 102 of the fourth modification, even if the content superimposing apparatus 140 does not store the feature amount of the reference points of a large amount of reference images such as 5000 books × 300 pages in the database, only necessary data is stored. Can be downloaded from the external search server 240.

  Further, in the case of transmitting the binary code feature quantity of the feature point of the input image to the external search server 240 in order to specify the necessary feature point database update data (that is, in order to specify the book), the external search server In order to ensure the accuracy of the search in the matching unit 42 of 240, the binary conversion unit 13 generates a large binary code, transmits this to the external search server 240, and updates the feature point database from the external search server 240. After the data is downloaded, the binary conversion unit 13 can convert the feature amount detected from the input image into a binary code having a size corresponding to the size of the binary code in the feature point database update data.

  In Modification 4, the binary code in the feature point database update data downloaded from the external search server 240 by the content superimposing device 140 can be configured to improve only the ability to identify the book. That is, unlike the case of identifying the target from any arbitrary image as in the external search server 240 of the modification 4, the number of identification targets is small when only the image of the page for each book is the target of identification. Therefore, a conversion matrix w used for binary conversion can be generated for each book by machine learning, and a binary code can be generated so that matching can be performed efficiently.

  At this time, the transformation matrix w can be downloaded from the external search server 240 to the content superimposing device 140 together with the feature point database update data of the book. Then, after downloading the feature point database update data of the book, the binary conversion unit 13 of the content superimposing apparatus 140 performs binary conversion on the feature amount of the feature point of the input image using the conversion matrix w downloaded therewith. With this configuration, the capacity of the feature point database 14 and the calculation cost of the corresponding reference image search device 10 or the content superimposing device 140 can be expected to be reduced as compared with the case where images of any arbitrary book pages are targeted for matching. .

  Note that when the book targeted for AR is changed, the feature point database update data downloaded from the external search server 240 to the feature point database 14 of the content superimposing device 140 cannot be used. In this case, the matching reference image cannot be detected by the matching unit 15. Therefore, when the matching unit 15 can no longer detect the corresponding reference image, the binary code of the feature amount of the feature point of the input image is transmitted again to the external search server 240 via the communication unit 31, and the external search server 240. Then, matching is performed to identify a new book, the content superimposing device 140 downloads new feature point database update data, and the feature point database 14 is updated.

  In the above-described embodiment and each modification, the feature point database is assigned an image identification number for each image, and the matching unit is included in one image detected as a corresponding reference image. The corresponding points were output to the correspondence calculation unit. However, the present invention is not limited to this. In the present invention, an image identification number may be assigned to each target to which content is added in the feature point database, with the target to which content is assigned as a reference image. That is, the reference image may not be one image, and a set of feature points included in one image may be used as the reference image.

  For example, when there is an image shown in FIG. 2, a set of feature points related to a mountain-shaped object (see FIG. 3) in this image is taken as one reference image, and a set of feature points related to a cloud-shaped object ( 3) may be used as another reference image. In this case, the feature point database is assigned different image identification numbers for the feature points related to the Yamagata object and the feature points related to the cloud object, and the matching unit receives a lot of votes as a corresponding reference image. Of the feature points related to the detected object (feature points assigned the same image identification number), the points corresponding to the feature points of the input image are output to the correspondence calculation unit as the corresponding points of the corresponding reference image. In this case, the content database also stores content for each target.

  As described above, the present invention has an effect that a reference image corresponding to an input image can be searched effectively even in an apparatus with limited resources, and an input image is obtained using feature points of the image. This is useful as a corresponding reference image search device that searches for a corresponding reference image.

DESCRIPTION OF SYMBOLS 10, 20 Corresponding reference image search apparatus 11 Image acquisition part 12 Feature quantity detection part 13 Binary conversion part 14 Feature point database 15 Matching part 16 Environment measurement part 21 Correspondence relation calculation part 22 Content conversion part 23 Content database 24 Superimposition part 31 Communication part 41 Communication Unit 42 Matching Unit 43 Feature Point Database 44 Content Database 101, 102 Content Superimposition System 110, 120, 130, 140 Content Superimposition Device 230, 240 External Search Server

Claims (15)

A corresponding reference image retrieval device for retrieving a reference image corresponding to an input image,
A feature quantity detection unit that extracts feature points from the input image and detects the feature quantities of the feature points;
A binary conversion unit that converts the feature amount detected by the feature amount detection unit into a binary code;
A feature point database storing feature amounts of feature points of a plurality of reference images in a binary code format;
By comparing the feature amount of the binary code of the input image converted by the binary conversion unit with the feature amount of the binary code of the plurality of reference images stored in the feature point database, the plurality of references A matching unit for detecting a reference image corresponding to the input image from the images;
A corresponding reference image retrieval apparatus comprising:   The corresponding reference image search device according to claim 1, wherein the binary conversion unit converts the feature quantity detected by the feature quantity detection unit into a binary code using a transformation matrix.   The corresponding reference image retrieval apparatus according to claim 2, wherein the transformation matrix is a sparse matrix.   The corresponding reference image search device according to claim 2, wherein the binary conversion unit can change the size of the binary code by changing the size of the conversion matrix.   5. The correspondence reference according to claim 1, wherein when there are a plurality of reference images corresponding to the input image, the matching unit detects a plurality of reference images. 6. Image search device. An environment measurement unit that measures an execution environment of the corresponding reference image search device;
The corresponding reference image search according to claim 2, wherein the binary conversion unit changes the size of the binary code by changing the size of the conversion matrix according to a measurement result by the environment measurement unit. apparatus. A content superimposing device comprising the corresponding reference image search device according to any one of claims 1 to 6, and superimposing a corresponding content on the input image,
A content database storing correspondence between the content and the reference image and the content;
A content extraction unit for extracting content corresponding to the reference image detected by the matching unit from the content database;
A superimposing unit that superimposes the content extracted by the content extracting unit on the input image;
A content superimposing apparatus comprising: The feature amount detection unit extracts a feature point including information on a position in the input image,
The feature point database stores information on the position of each feature point together with the feature amount of each feature point of a plurality of reference images.
The content database further stores the superimposed position of the content,
The content superimposing device is further extracted by the content extraction unit based on the relationship between the position of the feature point extracted by the feature amount detection unit and the position of the feature point stored in the feature point database. A content conversion unit that converts the superimposed position stored in the content database of the content,
The content superimposing apparatus according to claim 7, wherein the superimposing unit superimposes the content extracted by the content extracting unit on a superimposition position converted by the content converting unit in the input image. A content superposition system comprising a content superposition device and an external search server capable of communicating with the content superposition device,
The content superimposing device includes:
A feature quantity detection unit that extracts feature points from the input image and detects the feature quantities of the feature points;
A binary conversion unit that converts the feature amount detected by the feature amount detection unit into a binary code;
A content superimposing apparatus side communication unit that transmits the feature amount of the binary code of the input image converted by the binary conversion unit to the external search server;
With
The external search server
An external search server-side communication unit that receives the binary code feature quantity of the input image transmitted from the content superimposing device-side communication unit;
An external search server side feature point database storing feature amounts of feature points of a plurality of reference images in a binary code format;
Comparing the binary code feature quantity of the input image received by the external search server side communication unit with the binary code feature quantity of the plurality of reference images stored in the external search server side feature point database. An external search server side matching unit for detecting a reference image corresponding to the input image from the plurality of reference images,
A content superposition system characterized by comprising: The feature amount detection unit extracts a feature point including information on a position in the input image,
The feature point database stores information on the position of each feature point together with the feature amount of each feature point of a plurality of reference images.
The content superimposition system includes:
A content database storing a correspondence between the reference image stored in the feature point database and the content, and a superimposed position of the content;
The content corresponding to the reference image detected by the external search server side matching unit is extracted from the content database, and the feature point position extracted by the feature amount detection unit and the feature stored in the feature point database A content conversion unit that converts the superimposed position stored in the content database of the content extracted from the content database based on the relationship with the position of a point;
A superposition unit that superimposes the content extracted by the content conversion unit on the superposition position converted by the content conversion unit in the input image;
The content superposition system according to claim 9, further comprising: The content superimposing device includes:
A content superimposing apparatus-side feature point database storing feature quantities of feature points of a plurality of reference images in a binary code format;
By comparing the feature amount of the binary code of the input image converted by the binary conversion unit with the feature amount of the binary code stored in the content superimposing apparatus side feature point database, the plurality of reference images A content superimposing apparatus side matching unit for detecting a reference image corresponding to the input image from
With
The external search server side communication unit includes the feature amounts of the reference image detected by the external search server side matching unit among the feature amounts stored in the external search server side feature point database and the feature amount of the reference image related thereto. To the content superimposing device,
The content superimposing apparatus side communication unit receives the feature amount of the binary code transmitted from the external search server side communication unit,
The content superimposing apparatus side feature point database uses the characteristic amount of the binary code received by the content superimposing apparatus side communication unit as a feature amount of each of the feature points of the plurality of reference images. Item 14. The content superimposition system according to Item 9. A corresponding reference image retrieval method for retrieving a reference image corresponding to an input image in a corresponding reference image retrieval apparatus having a feature point database storing feature quantities of feature points of a plurality of reference images in a binary code format. And
A feature point extraction step for extracting feature points from the input image;
A feature amount detection step of detecting a feature amount of the feature point extracted in the feature point extraction step;
A binary code conversion step of converting the feature quantity detected in the feature quantity detection step into a binary code;
By comparing the feature amount of the binary code of the input image converted in the binary code conversion step with the feature amount of the binary code of each of the plurality of reference images stored in the feature point database, A matching step of detecting a reference image corresponding to the input image from a plurality of reference images;
And a corresponding reference image search method. Corresponding reference image retrieval apparatus having a feature point database storing feature quantities of feature points of a plurality of reference images in the form of binary code, and correspondence between reference images stored in the feature point database and contents A content superimposing method for superimposing corresponding content on an input image in a content superimposing apparatus comprising a stored content database,
A corresponding reference image search step for detecting a reference image corresponding to the input image in the corresponding reference image search method according to claim 12,
A content extraction step of extracting content corresponding to the reference image detected in the corresponding reference image search step from the content database;
A superimposing step of superimposing the content extracted by the content extraction unit on the input image;
The content superimposing method characterized by including.   A computer program for causing a computer to execute the corresponding reference image search method according to claim 12.   A computer program for causing a computer to execute the content superimposing method according to claim 13.

Images