JP2013109658A - Image retrieval device, image retrieval method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image retrieval device, an image retrieval method, and a program capable of retrieving an image area similar to a template image quickly with high accuracy.SOLUTION: An image retrieval device includes: first calculation means for calculating first similarity that indicates similarity between a retrieval target image and a template image at each retrieval position, while moving the retrieval positions by the template image to the retrieval target image; specifying means (S2) for specifying, for the first similarities within the predetermined number of highest ones, retrieval positions in which each of the predetermined number of first similarities has been calculated; second calculation means (S3) for calculating, for the retrieval positions, second similarity having calculation accuracy higher than that of the first similarity; and retrieval means (S4) for performing retrieval to the retrieval target image by the template image by using the second similarity as an initial value of a threshold.

Description

  The present invention relates to an image search apparatus, an image search method, and a program for searching an image area similar to a template image in a search target image.

  The full search method is most used as an image processing technique for searching an image region similar to a template image in a search target image. In this full search method, when the most similar T location (T is an integer of 1 or more) is detected, the search target image and the template are shifted at each position while shifting the template image on the search target image by one pixel interval. As a similarity measure for measuring the degree of similarity between images, for example, the absolute difference value sum Sum of Absolute Differences (SAD) is calculated, and template images in T search target images that give from the smallest SAD to the Tth smallest SAD, Alternatively, the position of a similar image of the template image is found.

  Here, the pixel value at the position (i, j) in the template image A is a (i, j) (1 ≦ i ≦ N, 1 ≦ j ≦ M), and the position (i, j) in the search target image B. The pixel values of b (i, j) (1 ≦ i ≦ X, 1 ≦ j ≦ Y, N ≦ X, M ≦ Y) are searched for the same position as the coordinates (1, 1) of the template image A as the search position. When the coordinate value of the target image B is (c, d) (1 ≦ c ≦ X−N + 1, 1 ≦ d ≦ Y−M + 1), the SAD at the search position (c, d) is expressed by the following equation (1). Can be sought.

…（１）

... (1)

In the full search method, T values are prepared as threshold values, and the SAD from the smallest value to the Tth smallest SAD up to the search position is changed from SAD _{min, 1} to SAD _{min, t.} If SAD is larger than SAD _{min, t,} which is the Tth smallest value, the detection is skipped as not being a detection candidate. If SAD _{min, t} or less, SAD _{min, t} is discarded _, and SAD at the detection position is A new threshold value is added, and the Tth smallest value is newly set as SAD _{min, t} after sorting among T pieces. After performing the above processing on all the pixels, the positions where T threshold values are calculated are output as detection positions. Moreover, according to the case, a result is output in order of a small threshold value.

  On the other hand, the Coarse-to-fine method has been proposed as one of the methods for reducing the calculation time in the full search method. This method is different from the full search method in which the template image is shifted by one pixel interval to calculate SAD. First, the entire search target image is searched at a coarse pixel interval to give the Tth smallest SAD from the smallest SAD. It is the smallest based on a coarse-to-fine search strategy, in which T positions are roughly estimated, and then the positions obtained by the search are searched with finer pixel intervals to estimate a more optimal position. The position of the template image in T search target images giving values from SAD to the Tth smallest SAD is found.

  In addition, as a technique for searching at a higher speed than the full search method, the similarity obtained by the same iterative calculation as the full search method such as SAD and Euclidean distance is applied as the similarity between the search target image and the template image. When the similarity in the middle of the calculation at the search position exceeds the threshold value obtained so far, the sequential similarity detection algorithm (SSDA method) that stops the calculation and moves to the next search position (for example, Non-Patent Document 1) Have been proposed). Hereinafter, the degree of similarity when the calculation is terminated when the threshold value is exceeded is referred to as “ssda distance”.

  As another technique for performing a search at a higher speed than the full search method, a Successive Elimination Algorithm (SEA method) has been proposed (see, for example, Non-Patent Document 2). The SEA method uses the triangular inequality | a + b | ≦ | a | + | b | to reduce the number of times to calculate SAD by determining whether it is necessary to calculate SAD at each search position, This is a technique for speeding up image search processing.

  FIG. 4 is a flowchart showing the flow of processing operations of the SEA method. As shown in FIG. 4, in the SEA method, sea (c, d), which is the sea distance at each search position, is calculated by the following equation (2) (step S31).

…（２）

... (2)

Next, it is determined whether or not the sea distance is larger than SAD _{min, t,} which is the Tth smallest SAD obtained in the search so far (step S32), and sea (c, d)> SAD _{min, If t} , the search position (c, d) is skipped and the next search position is started. On the other hand, if sea (c, d) ≦ SAD _{min, t} , SAD (c, d) is calculated at this search position (step S33), and the magnitude of SAD (c, d) with respect to SAD _{min, t} is determined. (Step S34). If SAD (c, d)> SAD _{min, t} , the process proceeds to the next search position. If SAD (c, d) ≦ SAD _{min, t} , the already stored SAD _{min, t} is discarded, and the SAD at the detection position is newly added as a threshold value, sorted among T, and the Tth A new value is newly set as SAD _{min, t,} and this search position is stored (step S35). This process is performed at all the search positions (steps S36 and S37), and after the search is completed up to the last search position, the smallest SAD _{min, t} value and the search position (c, d) from which the value is calculated As a search result (step S38). Thereby, the search position of the image area most similar to the template image is output. The output target at this time is not limited to the value of SAD _{min, t} and the search position (c, d), and _T search positions from which the final SAD _{min, 1} to SAD _{min, t} are obtained are shown. It is good also as a form output as a search result.

Here, when a similar image region is searched using SAD by the SEA method, when a value larger than SAD _{min, t} is obtained as sea (c, d), SAD (c, d)> SAD _The reason why the search position can be skipped will be described as _{min, t} .

  From the triangle inequality | x | + | y | ≧ | x + y |

…（３）
となる。したがって、

... (3)
It becomes. Therefore,

…（４）
であれば、

(4)
If,

…（５）
がｓｅａ距離であるため、

... (5)
Is the sea distance,

…（６）
となるので、当該検索位置でのＳＡＤを計算する必要がなく、スキップ可能となる。

... (6)
Therefore, it is not necessary to calculate the SAD at the search position, and skipping is possible.

  Further, a Multilevel Successive Elimination Algorithm (MSEA method) has been proposed in which an image is equally divided into a plurality of rectangles at a plurality of layers and a sea distance and a msea distance are calculated for each rectangle in a deeper layer until the threshold is exceeded. (For example, see Non-Patent Document 3).

  FIG. 5 is a flowchart showing the flow of processing operation of the MSEA method. As shown in FIG. 5, in the MSEA method, as in the SEA method, sea (c, d), which is the sea distance, is calculated at each search position (step S41).

Next, it is determined whether or not this sea distance is larger than SAD _{min, t,} which is the Tth smallest SAD obtained in the search so far (step S42), and sea (c, d)> SAD _{min, t.} If so, the search position (c, d) is skipped, and the process proceeds to the next search position. When sea (c, d) ≦ SAD _{min, t} , msea (c, d), which is the first stage msea distance at this search position, is calculated by the following equations (7) and (8) ( Step S43).

…（７）

... (7)

…（８）

(8)

  Here, P1 (1) = 0, P1 (2) = N / 2, P1 (3) = N, Q1 (1) = 0, Q1 (2) = N / 2, Q1 (3) = M. .

  The msea distance is a total value of psea distances, which are sea distances calculated individually in a rectangular area obtained by dividing the template image. In the MSEA method, the rectangular area is obtained by equally dividing the template image in the x-axis direction and the y-axis direction.

Next, it is determined whether or not the msea distance is greater than SAD _{min, t} (step S44). If msea (c, d)> SAD _{min, t} , the search position (c, d) is skipped and the next Move to the search position. If msea (c, d) <SAD _{min, t} , it is determined whether or not msea (c, d) at all stages set in advance has been calculated (step S45). If not all stages of msea (c, d) have been calculated, the process proceeds to step S43, and the msea distance of the next stage is calculated using the expressions (9) and (10) at this search position. .

…（９）

... (9)

…（１０）

(10)

  Here, P2 (1) = 0, P2 (2) = N / 4, P2 (3) = 2 × N / 4, P2 (4) = 3 × N / 4, P2 (5) = N, Q2 ( 1) = 0, Q2 (2) = N / 4, Q2 (3) = 2 × N / 4, Q2 (4) = 3 × N / 4, Q2 (5) = M.

The second stage msea distance is calculated using a rectangular area obtained by further dividing each rectangular area in the first stage. Subsequently, similarly, based on the magnitude relationship between this msea distance and SAD _{min, t} , it is determined whether to skip the search or to calculate the msea distance of the next stage (step S45).

If msea distance is calculated up to all preset stages and the msea distance obtained in the last stage is msea (c, d) <SAD _{min, t} , SAD (c, d) is calculated at this search position. Calculation is performed (step S46), and the magnitude of this SAD (c, d) with respect to SAD _{min, t} is determined (step S47).

If SAD (c, d)> SAD _{min, t} , the process moves to the next search position. When SAD (c, d) ≦ SAD _{min, t} , SAD _{min, t} is discarded, SAD at the detection position is newly added as a threshold value, and the Tth smallest value is newly sorted by sorting among T pieces. SAD _{min, t,} and this search position is stored (step S48).

This processing is performed at all the search positions (steps S49 and S50), and after the search is completed up to the last search position, the smallest SAD _{min, t} value and the search position (c, d) from which the value is calculated As a search result (step S51). Thereby, the search position of the image area most similar to the template image is output. The output target at this time is not limited to the value of SAD _{min, t} and the search position (c, d), and _T search positions from which the final SAD _{min, 1} to SAD _{min, t} are obtained are shown. It is good also as a form output as a search result.

Here, when a similar image region is searched using SAD by the MSEA method, when a value larger than SAD _{min, t} is obtained as msea (c, d), SAD (c, d)> SAD _The reason why the search position can be skipped will be described as _{min, t} .

  When a similar image region is searched using SAD by MSEA method, as in the case of using SAD by SEA method,

…（１１）
が導出される。

... (11)
Is derived.

…（１２）
がｍｓｅａ距離の第１段階であるため、

(12)
Is the first stage of the msea distance,

…（１３）
となり、第１段階（ｓｅａ距離）から順にチェックしていき、ＳＡＤ_{ｍｉｎ,ｔ}より大きな値が得られた時点で、ＳＡＤ（ｃ，ｄ）＞ＳＡＤ_{ｍｉｎ,ｔ}となり、検索位置をスキップできる。

... (13)
Thus, the check is performed in order from the first stage (sea distance), and when a value larger than SAD _{min, t} is obtained, SAD (c, d)> SAD _{min, t} , and the search position can be skipped.

FIG. 6 is a schematic diagram for explaining the image search processing operation in the above-described conventional technology. As shown in FIG. 6, in the full search method, the template image 51 is slid pixel by pixel on the search target image 50, and the similarity at that position is determined. When SAD is used as the similarity, if the SAD (c, d) at the search position (c, d) is smaller than the minimum SAD _{min, t} searched so far, SAD _{min, t} = SAD (c , D) and the search position (c, d) is stored. The processing is performed on all image regions of the search target image _, and a position (c, d) where SAD _{min, t} and SAD _{min, t} are taken is output.

Considering that SAD (c, d) has a large calculation processing amount (that is, processing is heavy), SEA has a small calculation processing amount before calculating SAD (c, d) at each search position (c, d). The amount of processing that avoids the SAD calculation by determining whether it is necessary to calculate the SAD using the MSEA method or the MSEA method, or skipping the calculation and moving to the next search position. Reduce search and perform faster searches. As a determination method at this time, the sea distance or msea distance obtained by simplified processing is used, and the above value is compared with SAD _min .

  On the other hand, for the SEA method and the MSEA method, the sea distance or the msea distance is used to detect a plurality of positions where the sea distance or the msea distance is a minimum or a similar value in the search target image, and only this position is detected. There has been proposed a Global Elimination Algorithm (GEA method) that reduces the number of SAD calculation processes by calculating the SAD and using the position where the minimum SAD is taken as a search result (see, for example, Non-Patent Document 4). In the GEA method, when outputting T results, T or more positions where the sea distance or msea distance is the minimum or similar value are detected, SAD is calculated only for the detected positions, and the smallest SAD To T positions that take values from to the Tth smallest SAD can be processed as search results.

  Also, in contrast to the SEA method and the MSEA method, the sea distance or the msea distance is used to detect the position where the sea distance or the msea distance is the smallest in the search target image, and the SAD is calculated at this position. An initial threshold estimation (ITE method) that reduces the number of SAD calculation processes by performing a search from the start position again as a threshold and using the position where the smallest SAD is taken as a search result has been proposed (for example, non-conversion). (See Patent Document 5).

D. Barnea and H. Silverman `` A class of algorithms for fast digital image registration '' IEEE Transaction on Computers, vol.C-21, no.2 pp.179-186, 1972 W. Li and E. Salari "Successive Elimination Algorithm for Motion Estimation" IEEE Transactions on Image Processing, vol. 4, no.1, pp. 105-107, January 1995 X.Q.Gao, C.J.Duanmu, and C.R.Zou `` A Multilevel Successive Elimination Algorithm for Block Matching Motion Estimation '' IEEE Transactions on Image Processing, vol.9, no.3, pp. 501-504, March 2000 YW Huang, SY Chien, BY Hsieh, and LG Chen `` Global Elimination Algorithm and Architecture Design for Fast Block Matching Motion Estimation '' IEEE Transactions on Circuits and Systems for Video Technology, vol.14, no.6, pp.898-907, June 2004 Minoru Mori and Kunio Kanno “Acceleration of template matching by adaptive segmentation and initial threshold estimation,” IEICE Transactions (D), vol.J94-D, No.5, pp.881-892

  However, the full search method always obtains T optimal positions that take the Tth smallest SAD in the search target image, but there is a problem that it takes time to calculate SAD.

  The Coarse-to-fine method can perform image search processing at a higher speed than the full search method, but there is a problem that there is no guarantee that an optimum position can be obtained because the search is performed with a rough pixel interval.

Similarly, the GEA method can perform image search processing at a higher speed than the full search method, but the position where the sea distance or the msea distance is small is not necessarily a position from SAD _{min, 1} to SAD _{min, t} . There is a problem that there is no guarantee that the optimum position can be obtained.

Further, in the ITE method, since processing for detecting one optimum place is assumed, when the SAD at the position where the sea distance or the msea distance is the smallest takes a value smaller than SAD _{min, t} , it is larger than the value. There is a problem that a position where the value SAD _{min, t} is taken cannot be obtained.

Further, the SSDA method compares the ssda distance, the SEA method the sea distance, and the MSEA method the msea distance with the SAD _{min, t} obtained by the previous search. By avoiding the censoring or SAD calculation, it is possible to search for an optimal similar position with less calculation processing. Therefore, the smaller the SAD value that is minimized, the more efficiently the processing amount is reduced.

However, since it is unclear where the search position where the SAD _{min, t} can be obtained correctly exists in the search target image, if SAD _{min, t} is obtained near the search start position, the calculation at the subsequent search position is performed. However, as the position where SAD _{min, t} is obtained is farther from the search start position, the number of times or the amount of reduction is reduced at the subsequent search position. Therefore, the degree to which SAD calculation can be aborted or skipped depends on the search target image and the template image, and there is a problem that it is not possible to realize a very high-speed image search process depending on the application example.

  The present invention has been made in view of the above circumstances, and provides an image search device, an image search method, and a program that can search an image region similar to a template image at high speed and high accuracy. Objective.

  In order to achieve the above object, the image search apparatus according to claim 1, the similarity between the search target image and the template image at each search position is moved while the search position of the template image is moved with respect to the search target image. A first calculating means for calculating a first similarity indicating a predetermined number of first similarities calculated from the first similarity calculated by the first calculating means in descending order of similarity. With respect to the similarity, the specifying means for specifying the search position where each of the predetermined number of first similarities is calculated, and the search position specified by the specifying means, based on the first similarity A second calculation unit that calculates a second similarity with high calculation accuracy; and the second similarity calculated by the second calculation unit is used as an initial value of a threshold value, and the test for the search target image is performed. A search unit for searching by plate image, characterized by comprising a.

  According to the image search device of claim 1, the first calculation unit moves the search position of the template image with respect to the search target image, and the similarity between the search target image and the template image at each search position. A first similarity indicating the degree is calculated. Further, in the present invention, the specific unit calculates the predetermined number of first similarities in order from the highest similarity among the first similarities calculated by the first calculating unit. The search position where each of the number of first similarities is calculated is specified. Further, in the present invention, the second calculation means calculates a second similarity with higher calculation accuracy than the first similarity for the search position specified by the specifying means. The predetermined number may be a plurality of numbers or only one.

  In the present invention, the search unit searches the template image for the search target image using the second similarity calculated by the second calculation unit as an initial threshold value.

  As described above, according to the image search device of the first aspect, the search position based on the template image with respect to the search target image is moved, and the first degree indicating the similarity between the search target image and the template image at each search position. A search in which the predetermined number of first similarities are calculated for a predetermined number of first similarities in order from the highest similarity among the calculated first similarities. A position is specified, a second similarity having a higher calculation accuracy than the first similarity is calculated for the specified search position, and an image search is performed using the calculated second similarity as an initial value of a threshold value. By using the first similarity that is lower in calculation accuracy than the second similarity and can be calculated at high speed, the calculation of the similarity is terminated at a faster stage from the start of the search, and less calculation processing and Kind by comparison number Or calculations degree, order to be able to judge whether to skip the search position can search for the image area similar to the template image at high speed with high accuracy, an effect that.

  In the image search device according to the present invention, as described in claim 2, the first similarity is at least one of a sea distance and an msea distance, and the second similarity is SAD. You may do it. Thereby, compared with the case where the similarity having a larger calculation load than the sea distance and the msea distance is used as the first similarity, an image region similar to the template image can be searched at high speed.

  Further, in the image search device according to the present invention, as described in claim 3, the first similarity may be mncc, and the second similarity may be NCC. This produces an effect that an image region similar to the template image can be searched at a higher speed than when the similarity having a calculation load larger than mncc is used as the first similarity.

  In the image search device according to the present invention, as described in claim 4, an integral image of the search target image and the template image, or an integral image of a square value of the search target image and the template image is calculated. The image calculation means may be further provided, and at least one of the first calculation means and the second calculation means may calculate the degree of similarity of the integral image calculated by the image calculation means. Thereby, compared with the case where the said integral image is not applied, there exists an effect that the image area | region similar to a template image can be searched at high speed.

  On the other hand, in order to achieve the above object, the image search method according to claim 5 is configured to move the search position of the template image with respect to the search target image, while moving the search target image and the template image at each search position. Among a first calculation step for calculating a first similarity degree indicating a similarity degree and the first similarity degree calculated in the first calculation step, a predetermined number of first numbers in order from the highest similarity degree. For the similarity, the specifying step for specifying the search position where each of the predetermined number of first similarities is calculated, and the search position specified by the specifying step, based on the first similarity A second calculation step for calculating a second similarity having a high calculation accuracy, and the second similarity calculated in the second calculation step as an initial value of a threshold value. And a, a search step of performing a search by the template image against the target image.

  Therefore, according to the image search method of the fifth aspect, since it operates in the same manner as the first aspect of the invention, it is similar to the template image at high speed and with high accuracy, similar to the first aspect of the invention. There is an effect that the image area can be searched.

  On the other hand, in order to achieve the above object, a program according to a sixth aspect causes a computer to function as the image retrieval apparatus according to any one of the first to fourth aspects.

  Therefore, according to the program of the sixth aspect, since the computer can be operated in the same manner as the image search apparatus of the present invention, an image similar to the template image at high speed and with high accuracy as in the image search apparatus. There is an effect that a region can be searched.

  According to the present invention, it is possible to retrieve an image region similar to a template image at high speed and with high accuracy.

It is a block diagram which shows the structure of the image search device which concerns on embodiment. It is a flowchart which shows the flow of operation | movement of the image search device which concerns on embodiment. It is a flowchart which shows the flow of the detailed operation | movement of the process in step S2 shown in FIG. It is a flowchart which shows the flow of the processing operation (SEA method) of the image search by a prior art. It is a flowchart which shows the flow of the processing operation (MSEA method) of the image search by a prior art. It is a schematic diagram with which it uses for description of the processing operation of the image search in a prior art.

  Hereinafter, an image search apparatus according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the present embodiment. In FIG. 1, reference numeral 100 denotes an image search device that is configured by a computer device and searches an image region similar to a template image in a search target image. The image search apparatus 100 includes an input unit 10, an initial threshold setting unit 20, an image search unit 30, and an output unit 40.

  The input unit 10 includes image information indicating a search target image (hereinafter referred to as search target image information) and image information indicating a template image that is an image to be searched on the search target image (hereinafter referred to as template image information). Enter. The input method at this time is arbitrary, and may be input from an external device by wired communication or wireless communication, or may be input from the recording medium by connecting the recording medium to the image search apparatus 100.

  The initial threshold setting unit 20 has a search target area (the same size as the template image indicated by the template image information input by the input unit 10) on the search target image indicated by the search target image information input by the input unit 10. Search position) is set, and the first similarity indicating the similarity between the search target image and the template image at each search position is calculated while moving the search position based on the template image.

  When calculating the first similarity, the initial threshold setting unit 20 moves the search position pixel by pixel from the coordinates (0, 0) of the upper left corner of the search target image, for example, and the search target image at each search position. A first similarity with the template image is calculated. Note that the start position of the search position at this time is not limited to the upper left corner of the search target image, but may be any position of the search target image such as the upper right corner, the lower left corner, or the lower right corner.

  The initial threshold value setting unit 20 calculates the similarity from the calculated first similarity to the value indicating the similarity from the value indicating the most similar to the value indicating the similarity in a predetermined order. Identify the location. In the present embodiment, an integer of 1 or more is applied as the predetermined order T.

  Then, the initial threshold setting unit 20 calculates the second similarity at the specified search position.

  Further, if T is not 1, the initial threshold setting unit 20 sorts the calculated T second similarities in descending order of similarity, and then searches the calculated second similarity values for an image search. To the unit 30. Note that the output second similarity value is used as an initial threshold value for various determination processes when the image search unit 30 performs a search. Below, in order to avoid complication, the case where the order | rank T is 2 or more is demonstrated.

The first similarity and the second similarity according to this embodiment satisfy either one of the following conditions (A) and (B).
(A) Calculation load: “first similarity” <“second similarity”
(B) Calculation accuracy: “first similarity” <“second similarity”

  The first similarity and the second similarity may satisfy both the above conditions (A) and (B). Hereinafter, in the present embodiment, description will be made assuming that the first similarity and the second similarity satisfy both the conditions (A) and (B).

  In the present embodiment, the above-described sea distance and msea distance are used as the first similarity (the sea distance is when the number of divisions is 0 (zero) in the msea distance), and the second similarity is described above. SAD is used.

  The first and second similarities mean that the smaller the value, the more similar. Therefore, the image region that is most similar in each similarity means an image region in which the sea distance, msea distance, or SAD between the search target image and the template image is minimized. Similarly, the T-th similar image region in each similarity means an image region in which the sea distance, msea distance or SAD between the search target image and the template image is the T-th smallest value.

  Further, when calculating the first similarity at each search position in the search target image, the initial threshold setting unit 20 generates an integrated image for the search target image and an integrated image for the template image, respectively. You may calculate using the integral image of.

  On the other hand, the image search unit 30 performs an image search with a template image for the search target image using the second similarity input from the initial threshold setting unit 20 as an initial threshold. At this time, the image search unit 30 sets the second similarity value as T initial threshold values from the smallest value to the Tth smallest value, and an image having the same size as the template image in the search target image. An area is cut out, and an image search is executed using a predetermined similarity in the cut out image area (search position).

For example, when the SSDA method is used as a search method, the ssda distance in the middle of calculating the similarity between the template image and the search position in the search target image, and the Tth similarity degree obtained so far are the highest. The threshold value (SAD _{min, t} ) is compared, and if the ssda distance is larger than the threshold value, the search position at that time is skipped. In this search position, when the calculation of the similarity is completed without skipping, the similarity obtained at the end (ssda distance = SAD) and the T-th similarity obtained so far are the highest. Only when the calculated similarity is smaller than the threshold, the threshold is discarded, the calculated similarity is newly added as a threshold, and the T thresholds are sorted to obtain the most similar The threshold from the highest degree to the threshold having the highest T-th similarity is newly set as T thresholds from SAD _{min, 1} to SAD _{min, t} .

Further, when the full search method is used as the search method, the SAD between the template image and the search position in the search target image is compared with a threshold (SAD _{min, t} ) having the highest T-th similarity, and the SAD If the value is larger than the threshold, the search position at that time is skipped. Otherwise, the threshold value (SAD _{min, t} ) having the highest T-th similarity is discarded, the calculated SAD is newly added as a threshold value, and the T threshold values are sorted to obtain the highest similarity threshold value. To the Tth threshold having the highest degree of similarity is newly set as T thresholds from SAD _{min, 1} to SAD _{min, t} . This process is repeated while shifting the search position on the search target image, and T searches corresponding to the T-th highest similarity value from the highest similarity level to the template image are performed. The position is searched for as T image areas similar to the template image in the search target image.

In addition, when the SEA method or the MSEA method is used as a search method, the similarity (sea distance or msea distance) at each division stage between the template image and the search position in the search target image is calculated, and the calculated similarity By comparing the threshold value (SAD _{min, t} when SAD is used) with the T-th highest similarity obtained in the previous search, the search position of the search target image at that time is compared with the template image. It is determined whether to calculate a detailed similarity (SAD). At this time, if the similarity at the final stage is higher than the threshold with the highest similarity of the similarities obtained in the previous search, the above detailed similarity is calculated. The threshold is discarded only when the similarity is smaller than the threshold having the highest T-th similarity obtained in the previous search, and the calculated detailed similarity is newly set as the threshold. After the addition, T threshold values are sorted, and a value from the highest similarity level to the T-th highest similarity level is newly set as SAD _{min, 1} to SAD _{min, t} . This process is repeated while shifting the search position on the search target image, and T search positions corresponding to the T-th highest similarity value from the value with the highest similar similarity to the template image are described. The search is performed as an image area most similar to the template image in the search target image.

  The output unit 40 outputs a search result by the image search unit 30, that is, information indicating an image region similar to the template image in the search target image. The information indicating the image area may be information indicating the position in the search target image of an image area similar to the template image in the search target image. The information indicating this position can exemplify, for example, the coordinate value of the upper left corner in the search target image of the image area.

  Next, the processing operation of the image search device 100 according to the present embodiment will be described with reference to FIG. FIG. 2 is a flowchart showing a flow of processing operations of the image search apparatus 100 according to the present embodiment.

  First, the input unit 10 inputs search target image information. Then, the input unit 10 inputs template image information that is a search query (step S1). After both pieces of image information are input, the initial threshold value setting unit 20 determines that the first similarity value is the most similar to the template image indicated by the template image information input by the input unit 10. The search positions that are T first similarities from the indicated value (minimum value) to the value indicating similarity at a predetermined rank T are specified (step S2).

  Next, the initial threshold value setting unit 20 obtains the second similarity at the specified T search positions, sorts the T second similarity, and starts from the second similarity having the smallest value. T values up to the second similarity with the Tth smallest value are output to the image search unit 30 (step S3).

  Subsequently, the image search unit 30 uses the T values input from the initial threshold value setting unit 20 as an initial threshold value, performs an image search process using the initial threshold value, and detects T image regions as similar parts ( Step S4).

  The output unit 40 outputs the similar part detected by the image search unit 30 as a search result (step S5).

  With the above processing operation, the image search device 100 according to the present embodiment can search T image regions similar to the template image in the search target image. Note that the search result output by the output unit 40 is not limited to outputting T image areas, and may output one image area most similar to the template image. Further, when outputting the image region, the output unit 40 displays the search target image on a display device (not shown), and T images similar to the template image on the displayed search target image. For example, the region may be displayed with a rectangular frame.

  Next, with reference to FIG. 3, the process by the initial threshold value calculation unit 20 executed in step S2 shown in FIG. 2 will be described in detail. First, the initial threshold value setting unit 20 calculates an integral value of pixel values at a high speed when calculating the first similarity and the second similarity, so that an integral image corresponding to each of the template image and the search target image is obtained. Create (step S20). For example, the integral value ai (i, j) at the coordinates (i, j) of the integral image ai with respect to the template image a is calculated by the equation (14).

…（１４）

... (14)

  However, a (0, 0) = a (0, 1) = a (1, 0) = 0.

  Similarly, the integral value bi (i, j) at the coordinates (i, j) of the integral image bi with respect to the search target image b is calculated by the equation (15).

…（１５）

... (15)

  However, b (0,0) = b (0,1) = b (1,0) = 0.

  Next, the initial threshold value setting unit 20 sets an area (search position) having the same size as the template image on the search target image, and moves the search position pixel by pixel from the coordinates (0, 0) of the upper left corner, for example. However, the first similarity between the search target image and the template image at each search position is calculated (step S21). This first similarity is a similarity for threshold search in the image search unit 30, and uses formula (2) indicating the sea distance, formula (7) indicating the msea distance, and the like. In the following steps, a case where the sea distance and the first stage msea distance are used will be described.

The initial threshold value setting unit 20 performs a process of calculating the sea distance at the search position (c, d) of the search target image. Subsequently, the initial threshold value setting unit 20 determines whether or not the sea distance is _{equal to} or greater than TH _{min, t} which is the Tth smallest threshold value obtained by the previous image search (step S22). In the present embodiment, each of the T threshold values from TH _{min, 1} to TH _{min, t} is set to a sufficiently large value as an initial value before starting execution of step S2. If the result of this determination is sea (c, d) ≧ TH _{min, t} , the initial threshold value setting unit 20 skips the search position (c, d), and has the search at all positions been completed? It is determined whether or not (step S26). If the result of determination in step S26 indicates that the search has not been completed at all positions, the initial threshold value setting unit 20 moves to the next search position (step S27).

On the other hand, if the determination result in step S22 is not sea (c, d) ≧ TH _{min, t} , the initial threshold setting unit 20 sets msea (c, d), which is the first stage msea distance at this search position ( 16) The calculation is performed according to the equation (step S23).

…（１６）

... (16)

  Here, the initial threshold value setting unit 20 uses the previously created integral images ai and bi in order to calculate each psea distance at high speed. Here, the psea distance of U1 (k) <x ≦ U1 (k + 1) and V1 (k) <y ≦ V1 (k + 1) is calculated by the equations (17), (18), and (19).

…（１７）

... (17)

…（１８）

... (18)

…（１９）

... (19)

  U1 (1),..., U2 (3) are the coordinates of the division position in the x-axis direction when obtaining the first stage msea distance. In addition, V1 (1),..., V1 (3) are also coordinates of the division position in the y-axis direction when obtaining the first stage msea distance, similarly to U1 (1),. .

Next, the initial threshold value setting unit 20 determines whether or not the calculated msea distance is greater than or equal to TH _{min, t} (step S24). When the determined result is msea (c, d) ≧ TH _{min, t} , the initial threshold value setting unit 20 skips the search position (c, d) and moves to the next search position (steps S26 and S27). . On the other hand, when the determined result is not msea (c, d) ≧ TH _{min, t} (when msea (c, d) <TH _{min, t} ), the initial threshold setting unit 20 sets TH _{min, t} to this msea ( c, d), T pieces from TH _{min, 1} to TH _{min, t} are sorted, and the Tth smallest threshold value from the minimum value is set again as TH _{min, 1} to TH _{min, t} .

  Further, the initial threshold value setting unit 20 stores the search position (c, d) at this time (step S25). The initial threshold value setting unit 20 executes such processing at all search positions in the search target image (steps S26 and S27). In this way, the search is completed up to the last search position in the search target image.

Next, details of the processing operation of the image search unit 30 (processing operation of step S4 in FIG. 2) will be described. As described above, in step S3, the initial threshold value setting unit 20 calculates SAD (c, d) at T search positions (c, d) corresponding to TH _{min, 1} to TH _{min, t.} Then, the obtained T SADs are sorted and output to the image search unit 30. In response to this, the image search unit 30 sets T values from SAD _{min, 1} to SAD _{min, t ,} which are the Tth smallest SAD from the smallest SAD, as the initial threshold values. The image search unit 30 uses the T values from SAD _{min, 1} to SAD _{min, t} as threshold values to execute an image search for the search target image at each search position.

The image search unit 30 performs an image search process using, for example, a known MSEA method (see FIG. 5). Here, if the image search unit 30 can reduce the amount of calculation processing by using SAD _{min, 1} to SAD _{min, t} as a threshold, a method other than the MSEA method is applied. Also good. For example, the image search process may be executed by the SEA method (see FIG. 4) or the SSDA method described above. Thereafter, the image search unit 30 outputs each search position (c, d) and SAD from which values from SAD _{min, 1} to SAD _{min, t} are obtained via the output unit 40 as a search result.

  In the above description, the sea distance and the msea distance are calculated at all search positions. However, unlike the full search method using SAD, the search can be executed with a very small calculation amount. In this case, the difference between the SEA method and the MSEA method is that the SAD calculation is not performed.

  Further, although the present embodiment uses both the sea distance and the msea distance when calculating the first similarity, it is not limited to this, and only one of the sea distance and the msea distance is used. Anyway. For example, when only the msea distance is used, the processes in steps S21 and S22 are omitted. On the other hand, for example, when only the sea distance is used, the processes of steps S21 and S22 are omitted, and the sea distance is used instead of the msea distance in steps S23 to S25.

As described above, in the image search device 100 according to the present embodiment, the initial threshold value setting unit 20 uses the first similarity for the search target image, and the first similarity is the smallest value. T search positions from T to the Tth smallest value are detected. Then, the initial threshold setting unit 20 calculates SAD, which is the second similarity used for the final image search determination, at these search positions. Next, the initial threshold value setting unit 20 sorts the T SADs to determine from the SAD having the smallest value to the SAD having the smallest value. Then, the initial threshold setting unit 20 outputs the determined T SADs to the image search unit 30. For the T SADs input from the initial threshold setting unit 20, the image search unit 30 sets the SAD having the smallest value from the smallest SAD as the initial threshold value from SAD _{min, 1} to SAD _{min, t.} Set and search within the search target image using the set initial threshold, and image regions from the highest similarity degree to the plurality of image areas to the T-th highest similarity degree, Search is performed as an image region similar to the template image in the search target image.

  As a result, when searching for a plurality of image regions similar to the template image in the search target image, the image search device 100 calculates the SAD with a smaller number of layers and the number of processes, or sets the search position at that time. It is possible to determine whether or not to skip, to cancel the calculation of the ssda distance with a smaller number of calculations, and to search an image area similar to the template image in the search target image at high speed.

  Further, the initial threshold setting unit 20 can calculate the sea distance and the msea distance at high speed by calculating the integrated value of the pixel values in the rectangular area using the integrated image.

  In the above description, the first stage of the sea distance and the msea distance is used as the first similarity, but only the sea distance or the distance after the second stage of the msea distance can be used.

  In addition, the initial threshold setting unit 20 calculates the first similarity while moving the search position pixel by pixel. However, to further increase the speed, the first similarity is calculated while moving at intervals of two or more pixels. It is also possible to calculate.

  In the above description, T> 1 is assumed, but by setting T = 1, it is possible to perform processing equivalent to the conventional ITE method.

  In the above description, the integral image is used when calculating the integral value of the pixel value. However, although the processing time naturally increases, a method of executing the pixel value addition process every time can be used.

In the above description, the MSEA method is used as the image search processing method of the image search unit 30, but the image search processing speed is increased by using the similarity value (SAD _{min, t,} etc.) as a threshold value. The present invention can be applied to any image search processing method (SSDA method, SEA method, etc.).

  As a result, the number of times the similarity calculation used to determine the final result of the image search such as SAD is terminated at an earlier stage of division than the conventional method, and the number of times this similarity calculation is omitted is increased, resulting in faster image search processing. Is possible. In addition, unlike the coarse-to-fine method, it is always possible to obtain an optimum result by executing an exhaustive search by the image search unit 30. Further, the image search apparatus 100 according to the present invention is not only used for image search, but also as a block matching motion prediction method. In the video sequence compression process, the image search apparatus 100 applies the most to the block currently being processed on two adjacent time axis screens. It can also be used to search for nearby blocks.

  In the above description, the initial threshold setting unit 20 uses the sea distance and the msea distance as the first similarity for determining the similarity between images, and the second similarity for determining the search result in the image search unit 30. SAD is used as the degree, but it is also possible to use a degree of similarity or a criterion different from these.

  In the following, for example, a normalized correlation (NCC) is used instead of SAD, and an NCC approximate value obtained from the Cauchy-Schwarz inequality is used instead of the sea distance and the msea distance.

  Unlike the case where the sea distance and the msea distance are used as the first similarity, when the NCC is used as the first similarity, the first similarity is more similar as the NCC value is larger. means.

  The value of NCC is given by the following equation (20).

…（２０）
ここで、

... (20)
here,

…（２１）

... (21)

…（２２）

... (22)

…（２３）
である。また、ここで‖Ａ‖は最初に一度計算すれば良く、また、‖Ｂ（ｃ，ｄ）‖は各検索箇所で１度計算すれば良く、それぞれ以後の計算は不要である。

... (23)
It is. Here, ‖A‖ may be calculated once at the beginning, and ‖B (c, d) ‖ may be calculated once at each search location, and subsequent calculations are not required.

  The Cauchy-Schwarz inequality is given by the following equation (24).

…（２４）

... (24)

  In this case, when the Cauchy-Schwarz inequality is applied to ψ (c, d), the following equation (25) is obtained.

…（２５）

... (25)

If the value obtained from the state divided in the q-th stage is mcc _levelq ,

…（２６）
となるので、ＭＳＥＡ法の場合と同様に、

... (26)
Therefore, as in the case of the MSEA method,

…（２７）
の関係が得られる。

... (27)
The relationship is obtained.

  Using ‖A‖ and ‖B (c, d) ‖ as the above value, mncc (c, d) at the search position is calculated by the following equation (28).

…（２８）

... (28)

Therefore, for example, instead of the sea distance and the msea distance in the search in the initial threshold setting unit 20 in the above-described example, the mncc is used from the 0th stage to the stage designated in advance, and the largest mncc to the Tth largest. T positions up to mncc are detected, T NCCs obtained at the detected positions are sorted to obtain the first to Tth NCC values in descending order, and each NCC value is stored in the image search unit 30. The image search unit 30 sets from the NCC having the largest value to the NCC having the largest value as the initial threshold value from NCC _{max, 1} to NCC _{max, t} , and using these T threshold values. The image search is performed with a smaller calculation amount by performing image search processing using NCC as the similarity for determining the image search result in the image search unit 30. It becomes possible.

  Also, by creating an integral image in advance at the time of SAD calculation, it is possible to calculate a square-value integral image in advance in the same way that addition (= integration) of values in each rectangular area can be calculated at high speed. If created, the addition (= integration) of square values in each rectangular area can be easily calculated. Specifically, the integral value ai2 (i, j) at the coordinates (i, j) of the integral value ai2 of the square value with respect to the template image a is calculated by the following equation (29).

…（２９）

... (29)

  However, ai2 (0,0) = ai2 (1,0) = ai2 (0,1) = 0.

  Similarly, the integral value bi2 (i, j) at the coordinate (i, j) of the integral image bi2 of the square value with respect to the search target image b is calculated by the following equation (30).

…（３０）

... (30)

  However, bi2 (0,0) = bi2 (1,0) = bi2 (0,1) = 0. The square integral value surrounded by each coordinate value is also obtained by the same method as the example of the integral image when SAD is used.

  In the present embodiment, the search target image and the template image are described as being grayscale images. However, the present invention is not limited to this, and the search target image and the template image may be color images. In this case, the first similarity and the second similarity described in the present embodiment are calculated for each primary color of R (red), G (green), and B (blue) in the color image information, and these similarities are calculated. Is used to calculate an initial threshold value at the time of image search.

  In addition, each configuration of the input unit 10, the initial threshold setting unit 20, the image search unit 30, and the output unit 40 included in the image search device 100 according to the present embodiment is realized by dedicated hardware. It may also be realized by a memory and a microprocessor. In addition, each configuration may be configured by a memory and a CPU (central processing unit), and the function may be realized by loading a program for realizing the function of each configuration into the memory and executing the program. .

  As described above, in the present invention, the template image and the search target image are similar to the Tth from the one that indicates that the similarity is the most similar using the first similarity with a small calculation load. A search position up to the one indicating that the image is detected, a second similarity used to determine a final image search processing result such as SAD is calculated at the specified search position, and the T second similarities are calculated. Sorting the degrees, and setting T second similarity degrees from the value with the highest degree of similarity to the value with the highest degree of similarity as the Tth as an initial threshold value used as a threshold value in various determination processes of image search, The normal search process is executed again from the search start position using the T initial threshold values. In the present embodiment, the first similarity is the sea distance, the msea distance, or the mncc that is an approximate value of NCC obtained from the Cauchy-Schwarz inequality, and the sea distance, the msea distance, and the mncc are further used as an integrated image. To calculate.

  As a result, the degree of similarity that is the same as the similarity value used for the determination of the search result such as SAD obtained at an appropriate search position or a value closer to the similarity value is used as a threshold value in the determination process is the highest degree of similarity. By executing an image search by setting an initial threshold value from the value to a value with the highest degree of similarity to the Tth, the calculation of the similarity degree such as SAD at the earlier stage from the start of the search is aborted, or fewer calculations are performed It is possible to determine whether the similarity is calculated by processing and the number of comparisons, or whether the search position is skipped, and an image area similar to the template image in the search target image can be searched at higher speed. In addition, by executing an exhaustive image search process using the initial threshold according to the present invention, it is possible to always obtain an optimal search result, unlike the coarse-to-fine method and the GEA method.

  It should be noted that a program for realizing the functions of the respective processing units of the image search apparatus 100 according to the present embodiment is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system. The image search process may be performed by executing. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within the scope not departing from the gist of the present invention. .

DESCRIPTION OF SYMBOLS 10 Input part 20 Initial threshold value setting part 30 Image search part 40 Output part 100 Image search apparatus

Claims (6)

First calculating means for calculating a first similarity indicating a similarity between the search target image and the template image at each search position while moving a search position of the template image with respect to the search target image;
Of the first similarities calculated by the first calculating means, the predetermined number of first similarities are calculated for a predetermined number of first similarities in descending order. Specifying means for specifying the search position,
Second calculation means for calculating a second similarity having higher calculation accuracy than the first similarity for the search position specified by the specifying means;
Search means for performing a search using the template image for the search target image, using the second similarity calculated by the second calculation means as an initial value of a threshold;
An image search apparatus comprising: The image search apparatus according to claim 1, wherein the first similarity is at least one of a sea distance and an msea distance, and the second similarity is SAD. The image search apparatus according to claim 1, wherein the first similarity is mncc, and the second similarity is NCC. An image calculating means for calculating an integral image of the search target image and the template image, or an integral image of a square value of the search target image and the template image;
The image according to any one of claims 1 to 3, wherein at least one of the first calculation unit and the second calculation unit calculates a similarity of the integral image calculated by the image calculation unit. Search device. A first calculation step of calculating a first similarity indicating a similarity between the search target image and the template image at each search position while moving the search position of the template image with respect to the search target image;
Among the first similarities calculated in the first calculating step, the predetermined number of first similarities are calculated for a predetermined number of first similarities in descending order. A specifying step for specifying the search position;
A second calculation step of calculating a second similarity having a higher calculation accuracy than the first similarity with respect to the search position specified in the specifying step;
A search step of performing a search using the template image for the search target image, using the second similarity calculated in the second calculation step as an initial value of a threshold;
An image search method characterized by comprising: Computer
The program for functioning as an image search device in any one of Claims 1 thru | or 4.

Images