JP4387889B2 - Template collation apparatus and method - Google Patents

Description

  The present invention relates to a template collation apparatus and method for detecting a signal similar to a known signal given as a template from an input signal that is input and taken in, and obtaining the position thereof, for example, when the input signal is sound / video Relates to a template collation apparatus and method suitable for CM / music detection technology from TV / radio broadcasting, etc., and in the case of images, for searching / position estimation of content such as a person or an object.

  A technique for searching for and finding a target object from an image is an elemental technique in a wide range of application fields such as searching from an image database, assigning an index to video data, and tracking an object. A typical technique is a template matching method (see Non-Patent Document 1, for example).

The template matching method is a technique for finding a partial region on an input image similar to an image (template) of a target object. A vector having the RGB value of each pixel in the template as an element is denoted as T. In the template matching method, a window having the same size as the template T is assigned to the input image, and the distance between the window and the template is calculated. As the distance, an absolute value distance (SAD: Sum of Absolute Differences) and a square distance (SSD: Sum of Square Differences) are widely used. When a vector whose elements are RGB values of each surface element of a window at a certain position (u, v) on the image is expressed by W ^{(u, v)} , the absolute value distance SAD is expressed by the following equation (1). It is.

Here, T ^p and W ^p are the p-th luminances of T and W ^{(u, v)} , respectively. The window is scanned on the input image (by changing u and v), and the distance is calculated, and the position where the distance is the minimum or less than the preset threshold is obtained.
The template matching method has a problem that the amount of calculation of the distance between the template and the window is large and the number of windows to be scanned is large. In order to reduce the amount of calculation, there are a method of reducing the amount of calculation of distance and a method of reducing the number of distance calculations.

  A typical method for reducing the amount of calculation of distance is a sequential residual test method (SSDA: Sequential Similarity Detection Algorithm) (see, for example, Non-Patent Document 2). In the sequential residual test method, a comparison is made with a threshold value at any time during distance calculation, and when the threshold value is exceeded, the processing is terminated. Chao et al. Effectively executes SSDA by appropriately changing the threshold value at the time of distance calculation according to the calculated number of pixels (see, for example, Non-Patent Document 3).

A typical method for reducing the number of distance calculations is a two-step template matching method (see, for example, Non-Patent Document 4). In the two-stage template, a sub-template that is a part of the template is used to calculate the lower limit value of the distance in each window, and only the windows whose distance lower limit value is equal to or less than the threshold value are collated with all template rest. Li et al. Calculate a region that does not require matching by applying a triangular inequality to the vector size | T | of the template T and the size | W | of each window vector W (non-patent document). 5).
Lee et al. Calculate an area that does not need to be collated efficiently by calculating a distance with a low-resolution image and calculating a distance lower limit value with a high-resolution image from the distance (for example, Non-Patent Document 6). reference). However, most of the above methods determine whether or not reduction is possible based on the difference between the already-matched distance between the window and the template and the threshold value.

On the other hand, the active search method we have developed so far calculates the upper limit value of similar values in a neighboring window using the result of matching in a certain window, and omits matching when the upper limit value falls below the threshold. Therefore, unnecessary verification can be omitted very efficiently (see, for example, Non-Patent Documents 7 and 8 and Patent Documents 1 and 2).
However, since the histogram is used for the feature, shape information and color distance information are lost, so it is difficult to apply when the difference in fine shape or positioning system is important, and different windows with similar color histograms It was also possible to falsely detect.

In addition, we have developed a Skipping Template Matching (STM) method that uses the similarity of sub-templates in a template to efficiently perform matching (see, for example, Non-Patent Documents 9 and 10).
According to the technology disclosed herein, when the similarity between sub-templates generated from a template of a target object is often high, the matching operation can be omitted more effectively than the conventional method described above. .

However, when the object has a three-dimensional shape, the image of the object photographed by the camera differs depending on the positional relationship between the camera and the object, so that the appearance (the direction and size of the object) differs. Need to prepare. There is a technique for searching for images of a large number of objects having different sizes and orientations as a technique for searching at high speed even when the images of the holiday are different. In order to reduce the increase in the amount of calculation due to the use of a large number of templates, a method of compressing the number of dimensions of the template using a partial empty method has been proposed (for example, see Non-Patent Document 11). However, the computational cost of dimensional compression is large, and further speedup is expected.
Onoe (edit): “Image processing handbook, Shosodo (1987) DI Barnea and HF Silverman: A class of algorithms for fast digital image registration ¥ ", IEEE Trans. On Comput ・, C-21, 2, pp.179-186 (1972) HC Chao, YPHung and WL Hwang: Adaptive early jump-out technique for fast motion estimation in video coding ", Proc. 0f ICPR1996, Vol.2, pp.864-868 (1996) A, Hatabu, T, Miyazaki and I. Kuroda: `` Optimization of decision-timing for early termination of SDDA-based block matching '', Proc. Of 1CME 2003, vol.2, pp.821-824 (2003) GJ, Vanderbrug and A. Rosenfeld: "Two stage template matching", IEEE Trans. On Comput, C-26, 4, pp. 384-393, 1 977 W. Li and E, Salari: '' Successive elimination algorithm for motion estimation '' FIEEE Trans. On Image Processing, 4.1, pp. 105-107 (1995) VV, Vinod and H. Murase: '' Focused color intersection with efficient searching for object extraction “, Pattern Recognition, 30, 10pp. 1787-1797 (1997) T. Kawanishi, H. Murase, S. Takagi and M. Werner: j 'Dynamic Active Search for quick object detection with pantilt-zoom camera ", Proc. Of ICIP 2001, Vol. 31 pp. 716-719 (2001) Kawanishi, Kurosumi, Sugano, Takagi: “High-speed template matching method using residual information of partial region of reference image”, FIT2003 Information Science and Technology Forum Information Letters, 2, pp.175-176 (2003) T, Kawanishi, T. Kurozumi, S. Takagi and K, Kashino: “Skipping Template Matching guaranteeing same accuracy as exhaustive search”, Proc. Of the Fifth Intl. Conf. Of Advances on Pattern Recognition, pp. 209-212 (2003 ) Murase, SKNayer: "3D object recognition by 2D matching-parametric eigenspace method-" Shinsei Gakkai (D-II), J77-D-II, 11, pp.2179-2187 (1994) Japanese Patent Laid-Open No. 9330404 JP 2002-31343 A

  The present invention has been made in view of the above circumstances, and even in the case where a histogram is not used, while guaranteeing the same collation accuracy as the entire collation, the sub-template which is a part of the template and a part of a certain local region. By using the result of matching, a template matching apparatus and method are provided that eliminates matching between a template and its local region, as well as a local region in the vicinity thereof, as much as possible, and further speeds up the template matching described above. For the purpose.

  In addition, as described above, a large number of templates are required for recognition of a three-dimensional object, and if the STM method is applied to each template independently, the calculation time increases in proportion to the number of templates. Templates having similar imaging parameters such as the distance to the camera are expected to have similarities. Therefore, another object of the present invention is to provide a template collation apparatus and method that can reduce the number of collations by utilizing the similarity between templates.

  In order to solve the above-described problem, the present invention is a template collation apparatus that detects a signal similar to a known signal given as a template from an input signal that is inputted and taken in, and obtains the position thereof. When calculating a distance value between a sub-template generation unit that generates sub-templates of different sizes, the input signal, and sub-templates sequentially generated by the sub-template generation unit, and confirming that they are similar, And a matching operation unit that additionally calculates a distance value with respect to the entire template.

  Further, in the present invention, the matching calculation unit calculates a distance value between all sub-templates having different sizes in stages and any partial area having the same size as the sub-template on the template. A calculation unit, a local region setting unit for setting any local region having the same size as the template on the input signal, a local region order setting unit for rearranging the order of the local regions, and the local region order setting unit One local region is extracted in the order set in step (3) and set as a candidate local region. The lower limit of the distance between the candidate local region and the template is compared with a predefined threshold value, and the candidate local region to be matched next is selected. From the plurality of subtemplates obtained by the candidate local region selection unit and the stepwise subtemplate generation unit, the size is reduced. A sub template selection unit that selects one sub template in order, a partial region of the candidate local region obtained from the candidate local region selection unit, and a sub template obtained by the sub template selection unit are sequentially collated. When the partial distance value calculation unit that calculates the partial distance value and the partial distance value obtained by the partial distance value calculation unit are similar, the distance between the remaining area of the template and the remaining area of the candidate local area When the distance value obtained by the additional distance value calculation unit for obtaining the distance value between the template and the candidate local region is less than the target threshold value, the detection result And the detection result update unit for updating the threshold, the self-distance value calculation unit, and the respective distance values calculated and output by the partial distance value calculation unit, A candidate local region pruning unit that deletes a local region candidate that is unlikely to have a signal similar to the template by updating a distance lower limit value of the local region in the vicinity of the region. .

  Further, the present invention is a template collation device that detects a signal similar to a known signal given as a template from an input signal that is inputted and taken in, and obtains the position thereof. A sub-template of a predetermined size is obtained from the template. A plurality of sub-template generation units, a reference sub-template determination unit that determines, for each template, a reference sub-template to be collated from the sub-template, and templates having similar characteristics among the templates are grouped, and A matching calculation unit that calculates in advance a self-distance value of a sub-template straddling the template and reflects the result of matching with the certain reference sub-template in the group.

  Further, in the present invention, the matching calculation unit is a self-distance value calculation unit that calculates all distance values between the reference sub template determined by the reference sub template determination unit and a sub template in the same group across the template. And a local region setting unit that sets every local region of the same size as the template on the input signal, a local region order setting unit that rearranges the order of the local regions, and the reference subtemplate determination unit A reference sub-template selection unit that sequentially extracts reference sub-templates from a set of reference sub-templates, a candidate local region selection unit that extracts one local region in the order set by the local region order setting unit and serves as a candidate local region, The distance between the candidate local region selected by the candidate local region selector and the template A distance lower limit evaluation unit that compares a limit value with a predefined threshold, a part of a candidate local region obtained from the candidate local region selection unit, and a reference subtemplate selected by the reference subtemplate selection unit If the partial distance value obtained by the partial distance value calculation unit and the partial distance value calculation unit that collate and calculate the partial distance value are similar, the remaining area of the template and the remaining area of the candidate local area When the distance value obtained by the additional distance value calculation unit calculating the distance value between the template and the candidate local region is less than the target threshold value And using the respective distance values calculated by the detection result update unit for updating the detection result and the threshold, the self-distance value calculation unit, and the partial distance value calculation unit, Characterized by comprising a candidate local region pruning unit to remove the local region candidate similar signal is not likely to be present in the template by updating the distance limit value of the local region.

  Further, in the present invention, the partial distance value calculation unit and the additional distance value calculation unit sequentially calculate the distance value, and it is determined that there is no need for calculation approximately or that calculation is not necessary. The calculation is terminated at the time.

  In the present invention, the sub template generation unit may generate a sub template in which the maximum value of the self distance value calculated by the self distance value calculation unit is minimized.

  In the present invention, the reference template determining unit may determine a sub template having a minimum maximum distance value between all the sub templates in the group as a reference sub template.

  In the present invention, the local region order setting unit rearranges the order at a predetermined interval.

  In the present invention, the local region order setting unit may perform rearrangement at any time based on the distribution of the distance values so as to have an appropriate interval.

  Further, in the present invention, the local region order setting unit arranges the positions of the centers of pyramid images obtained by recursively dividing the input signal in order from a large image, and uses the order for extracting the candidate local regions. .

  In the present invention, the local region order setting unit expresses the local region candidate order as a binary number, bit-inverts a value lower than the highest one, and rearranges them in ascending or descending order, The candidate local regions are used in the order of extraction.

  In the present invention, a one-dimensional sound signal is used for the input signal and the template.

  In the present invention, a one-dimensional video signal is used for the input signal and the template.

  In the present invention, a two-dimensional image signal is used for the input signal and the template.

  The present invention also relates to a template matching method for detecting a signal similar to a known signal given as a template from an input signal that is input and fetched, and obtaining the position thereof. A first step of generating a template; a second value for calculating a distance value between the input signal and sequentially generated sub-templates; These steps are included.

  Also, in the present invention, the second step is a sub-calculation for calculating a self-distance value with respect to all sub-templates having different sizes in a stepwise manner with any partial region having the same size as the sub-template on the template. A sub-step for setting all local regions of the same size as the template on the input signal, a sub-step for rearranging the order of the local regions, and extracting one local region in the set order and selecting candidates As a result of comparing the lower limit value of the distance between the candidate local region and the template with a predefined threshold as a distance area, a sub-step of selecting a candidate local region to be matched next, from the plurality of sub-templates, A sub-step of selecting one sub-template in ascending order of size, and the selected candidate A sub-step of calculating a partial distance value by sequentially comparing a partial area of the predetermined area and the selected sub-template, and if the obtained partial distance value is similar, the remaining area of the template Adding a distance with the remaining area of the candidate local area, substep for obtaining a distance value between the template and the candidate local area, and when the obtained distance value is below a target threshold value, The template is obtained by updating a distance lower limit value of a local region in the vicinity of the candidate local region using the detection result and a sub-step of updating the threshold, and the calculated and output self-distance value and partial distance value. And a sub-step of deleting a local region candidate that is unlikely to have a signal similar to.

Further, the present invention is a template matching method for detecting a signal similar to a known signal given as a template from an input signal that is inputted and taken in, and obtaining the position thereof. A sub-template of a predetermined size is obtained from the template. A first step of generating a plurality of steps, a second step of determining for each template a reference sub-template to be matched from the sub-template, and templates having similar characteristics among the templates, and grouping the templates within the group A third step of calculating in advance the self-distance value of the sub-template straddling the template and reflecting the result collated with the certain reference sub-template in the group;
It is characterized by having.

  In the present invention, the third step includes a sub-step of calculating all self-distance values between the determined reference sub-template and a sub-template in the same group across the template, and the input signal. A sub-step of setting any local region having the same size as the template, a sub-step of rearranging the order of the local regions, and a sub-step of sequentially selecting a reference sub-template from the set of the determined reference sub-templates; A sub-step of selecting candidate local regions in the set order; a sub-step of comparing a distance lower limit value between the selected candidate local region and the template with a predefined threshold; and the selected candidate local region Are compared with the selected reference sub-template to obtain a partial distance value. When the sub-step to calculate and the partial distance value are similar, add the distance between the remaining area of the template and the remaining area of the candidate local area, and the distance value between the template and the candidate local area A sub-step of calculating the detection result and the threshold when the distance value between the obtained template and the candidate local region is lower than a target threshold, and the calculated self A sub-step of deleting a local region candidate that is unlikely to have a signal similar to the template by updating a distance lower limit value of a local region near the candidate local region using a distance value and a partial distance value; and It is characterized by having.

  According to the present invention, by generating sub-templates of different sizes from the template, performing the collating operation in order from the smallest one, and collating with the sub-template having a large size step by step when it is determined that the distance value is close, Compared to the two-step template matching method, SSDA method, and AEJO method in the prior art, when the distance lower limit value exceeds the threshold as a result of matching a certain area, it is possible to omit not only the area but also the surrounding matching. Therefore, the target signal can be searched at high speed.

  Compared with the active search method that uses histograms for features, matching is performed including shape information and color distance information, so it is greatly affected by differences in fine shapes and changes in brightness. If the accuracy is good. This is particularly effective when there are many low-frequency components in the features in the template.

  Further, according to the present invention, similar templates are grouped together, and the self-distance value of the sub-template straddling the template within the group is calculated in advance, so that the result of matching with a certain reference sub-template The calculation is omitted by the principle of triangular inequality. For this reason, compared to the SSDA method, the template matching method, or the subspace method, if the distance lower limit value exceeds the threshold value as a result of matching a certain region, not only the region but also the surrounding matching can be omitted. The number of signals can be reduced, and the target signal can be searched at high speed. In particular, by grouping templates, the distance lower limit value can be calculated even across the templates, and the conventional STM method can be speeded up. Since a number of templates are collated at the same time, a remarkable effect can be obtained in searching for a three-dimensional object having a large number of templates.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment to which the present invention is applied.
The template collation apparatus of the present invention includes a template input unit 1, a sub template generation unit 2, a stepwise sub template generation unit 3, a self-distance value calculation unit 4, an input signal capture unit 5, and a local region setting unit 6. A local region order setting unit 7, a candidate local region selection unit 8, a sub template selection unit 9, a partial distance value calculation unit 10, an additional distance value calculation unit 11, a detection result update unit 12, and a local region The pruning unit 13 is configured to receive a template and an input signal, detect whether or not the template is included in the input signal, and output the position and distance value when the template is included.
In FIG. 1, the solid line arrows indicate the flow of processing, the dotted line arrows indicate the data flow, and each functional block is configured by software or hardware, or by a combination of software and hardware. It can be realized. FIG. 2 shows the processing procedure, FIG. 3 shows a template image, and FIG. 4 shows an example of an input image.

Here, the template and a part of the area in the input signal are evaluated by the distance value. First, a method for finding the most similar local region by giving an appropriate threshold value and updating the minimum value dmin found during the search as a threshold value will be described. Of course, it is easy to configure not to update the threshold value, but to record all the values below the threshold value and return the result.
Although a two-dimensional example is used here, it is obvious that a one-dimensional signal can be handled in exactly the same way by considering a one-dimensional input signal composed of F elements as a two-dimensional 1 × F. It is possible to handle signals of higher dimensions in exactly the same way. In addition to images, the present invention can also be applied to multidimensional vector sequences such as weather data, map data, and gene data.

Specifically, the template input unit 1 captures in real time a part of a one-frame image cut out from a two-dimensional image file given in various image formats or a moving image file given in various formats. An object region is extracted from a part of an image of one frame using a window function or the like, and sent to the sub-template generation unit 2 (S21). For example, the window function of an object can be raised such as a rectangular shape that extracts only the inside, a circular shape that extracts only the inside, or a shape that matches the shape of the object that extracts only the inside. It is done. Here, an example of a rectangular shape will be described.
This extracted signal is called a template. The template is represented by a vector having the pixel value as an element. This template is expressed as T. The sub template generation unit 2 generates a sub template R therein from the template T received from the template input unit 1 (S22). Here, for example, the upper left region or the lower right region of the template is used. Further, when searching for a single template from a large number of input images, it is possible to generate a sub template in which the maximum value of the self-distance value in the self-distance value calculation unit 4 is minimized.

Stepwise sub template generating unit 3 generates a plurality of phased sub-template R ^q of different sizes for the sub-template of acceptance from the sub-template generating unit 2 and the maximum value R ^Q. At this time, as shown in Figure, a number of sub-templates R ^O expanding ^stepwise, ..., R ^Q is obtained. The sub template at position _{(i, j)} is expressed as R ^q _{(i, j)} with the upper left corner of the template as the origin.
When the sub-template generation unit 2 uses the upper left region of the template as a sub template, this sub template can be expressed as R _{(0, 0)} . In this case, the stepwise subtemplate generation unit 3 also generates a stepwise subtemplate R ^q _(0,0) that expands from the upper left region of the subtemplate R _(0,0) , that is, the origin of the template (S23). On the other hand, if the sub-template generation unit 2 generates the lower right region of the template as a sub template, the stepwise sub template is generated based on the lower left of the template. All these sub-templates are sent to the self-distance value calculation unit 4 (S24).

Self distance value calculation unit 4, a gradual subphase obtained by the template generation unit 3 specific sub-template R q (i, j), stepwise sub-template R q (i, j) on the template with the same size A distance value (or similarity value) with the template partial region T q (k, l) is obtained for all graded sub-templates having different sizes. If the sub-template is smaller than the template size by (m, n), a large number of template partial areas permitting overlapping as shown in FIG. 6 are obtained. The total number is mxn per sub-template. Template The absolute distance, square distance, etc. can be used as the distance scale of this method. As a similar scale, for example, “Adachi et al.“ SSDA conversion using KL expansion of normalized cross-correlation template matching ”, IEICE The normalized cross-correlation values shown in the report PRMU96-173, pp. 23-30, 1997. can also be used. In the following example, the absolute value distance shown in Expression (2) is used.

Next, the distances d (R ^q _{(k, l, k)} between all the template partial regions T ^q _{(k, l)} whose sizes correspond to the sub templates R ^q _{(i, j)} received from the stepwise sub-template generation unit 3 _{. l)} , T ^q _{(k, l)} ) in advance step by step. Further, the distance d ( ^Rq _{(k, l)} , ^Tq _{(k, l)} ) and the maximum value _{dmax of} the distance d ( ^Rq _{(i, j)} , ^Tq _{(k, l)} ) are locally determined. The data is sent to the area pruning unit 13 (S31).

On the other hand, the input signal input unit 5 is an image of one frame cut out from a two-dimensional image file given in any of various image formats or a moving image file given in any of various formats. Part or all of the image is received in part or all in one frame image captured in real time, and the obtained vector S of pixel values is sent to the local region setting unit 6 (S25).
The local region setting unit 6 creates a set of all local regions to be collated in the input image received from the input signal capturing unit 5, and sends them to the local region order setting unit 7 (S26). The size of the local area is the same as that of the template T. A local region S (u, v) having the same size as the template T located at the coordinates (u, v) on the input image S is described. When the set of local regions is W = {S (u, v) | u, v}, the search by the template match sig method minimizes the distance d (T, S (u, v)) from W. This is a technique for finding a position S (u, v). The lower bound value bound (u, v) is introduced for all S (u, v) in W. The initial value of bound (u, v) for all (u, v) is 0. bound (u, v) = 0.

The local region order setting unit 8 sets the order of local regions included in W (the order in which the input image is scanned) (S27). For example, when the sub-template generation unit 2 takes out and collates the upper left area of the template as a sub template, scanning is performed in the order of “from lower right to upper left”.
Conversely, when collating using the lower right area of the template as a sub template, scanning is performed in the order of “from upper left to lower right”. In the case of the upper left template, it is possible to omit the local region shifted in the upper left direction, and in the case of the lower right sub template, it is possible to omit the local region shifted in the lower right direction. Because it does. On the contrary, the effect of omitting the search is not seen when scanning.

  On the other hand, in order to increase the speed, it is more effective to sequentially select the local regions that have been rearranged so as to have appropriate intervals. This method includes a method of rearranging as needed so as to have an appropriate interval based on the distribution of distance values, and a method in which the positions of the centers of pyramid images obtained by recursively dividing the input signal are arranged in order from the largest image. The method used for the local area candidate order and the continuous local area candidate order (upper left to lower right, etc.) are expressed in binary numbers, the values lower than the highest 1 are pit-inverted and rearranged in ascending order. There are methods. The W thus created is sent to the candidate local region selection unit 8.

The candidate local region selection unit 8 selects a candidate local region having a template from the set of local regions W received from the local region order setting unit 7, and sends it to the sub template selection 9 (S28). Then, the first S (u, v) is extracted from W. Here, if bound (u, v)> d _min , the candidate local region selection operation by the candidate local region selection unit 8 is executed again. If bound (u, v) ≦ d _min , (u, v) is sent to the sub-template selection unit 9 (S29). If there is no local area included in W, the search is terminated.
The sub template selection unit 9 selects the sub templates R ^q _{(i, j)} in ascending order from the sub template groups having different sizes obtained by the stepwise sub template generation unit 3 and sends them to the partial distance value calculation unit 10. When the partial distance value calculation unit 10 receives the position (u, v) of the candidate local region from the candidate local region selection unit 8, the partial distance value calculation unit 10 extracts the minimum subtemplate of the stepwise subtemplate, and the partial distance value calculation unit together with (uv) 10 to send. If these values are returned from the partial distance value calculation unit 9 as a possible target signal, the next largest sub-template R ^{q + 1} _{(i, j)} is selected.

The partial distance value calculation unit 10 receives the sub template R ^q _{(i, j)} received from the sub template selection unit 9 and the R ^q _(i ) of the candidate local region S (u, v) received from the candidate local region selection unit 9. _{, J)} and a distance d (R ^q _{(i, j)} S _{(i, j)} (u, v) between a part of the region S _{(i, j)} (u, v) having the same position and size. ). In the middle of this distance calculation, the pixels for which the distance calculation has been completed monotonously increase, so that the distance value also monotonously increases. Therefore, the partial distance value calculation unit 10 approximately evaluates whether or not the distance value eventually exceeds the threshold value during the calculation, and determines whether or not to terminate the calculation (S30). With this configuration, it is possible to use the SSDA method or the AEJO method that omits unnecessary distance calculation.

Here, d _min + d _max is used as the threshold value. When the distance value exceeds d _min + d _max , the partial distance value calculation unit 10 terminates the calculation because the target object is not included in this partial local region and the peripheral partial local region. Alternatively, when the distance value exceeds d _min after all the distance calculations are completed, similarly, the target object is not included in this local region. When the distance value is equal to or greater than d _min , the partial distance value calculation unit 10 sends the distance value d to the local region pruning unit 13 in order to further investigate whether or not an object is included in the surrounding region.

On the other hand, if the distance value does not exceed d _min after all calculations are completed, the object may be the target object, so the sub-template selection unit 9 is executed again to calculate the distance in a larger area, Next, the next larger sub-template R ^{q + 1} _{(i, j)} is selected. When the distance value d (R ^Q _{(i, j),} S ^Q _{(i, j)} (u, v) does not exceed d _min after the distance calculation of the maximum sub-template R ^Q _{(i, j)} is completed, The distance value is sent to the additional distance value calculation unit 11.

The additional distance value calculation unit 11 adds the distance between the remaining part of the template and the remaining part of the local region to the distance value of the region calculated by the partial distance value calculation unit 10, thereby The distance between them is obtained (S32).
Here, if the region of the template is T and the region not included in the sub template R ^Q _{(i, j)} is ~ R ^Q _{(i, j)} , the template T to the sub template R ^Q _{(i, j)} The area obtained by subtracting the area is T∩ ~ R ^Q _{(i, j)} . Similarly, a region obtained by subtracting the partial local region S ^Q _{(i, j)} (u, v) from the local region S (u, v) is expressed by the following equation (3).

That is, the partial distance value d (T ^Q _{(i, j)} (u, v), ~ R ^Q _{(i, j)} ) obtained by the partial distance calculation unit 10 is added to the distance d (T∩ ~ By adding R ^Q _{(i, j),} S (u, v) ∩ ~ S ^Q _{(i, j)} (u, v), the distance d (T, S (u, v) between the template and the candidate local region That is, the following equation (4) is obtained.

Here too, the SSDA method or the AEJO method is used as in the partial distance value calculation unit 11. If the result of calculating the distance value sequentially exceeds d _min in the middle, it is not the target object, so the distance value calculation is interrupted, and the local area selection process by the candidate local area selection unit 9 is executed. When the calculation is performed up to the last drawing, and the distance d between the template and the local area does not exceed d _min , d is sent to the detection result update unit 12.
The detection result update unit 12 updates the minimum value d _min with the distance value d obtained from the additional distance value calculation unit 11, records the position (uv), and executes candidate local region selection by the candidate local region selection unit 8. (S33).

On the other hand, the local region pruning unit 13 executes the lower limit value of the distance value between the template and the surrounding local region from the distance relationship between the obtained stepwise subtemplate and a part of the local region. In the local region that has obtained a large lower limit value, the candidate local region selection unit 8 can omit collation in the local region, so that the search can be speeded up without sacrificing accuracy. The principle is shown below.
The distance at which the AEJO method can be used is obtained by sequentially adding the distances for each pixel. Therefore, when the number of pixels to be compared increases, the distance value also increases monotonously. For this reason, the distance between the sub template R ^q _{(i, j)} as shown in FIG. 7 and a part S ^q _{(i, j)} (u, v) of the local region having the same size as the sub template, The relationship of the following formula (5) is established between the template and the distance d (T, S (u, v)) between the local regions.

That is, if the sub-templates do not match as a result of matching, the entire template does not match. Furthermore, when the principle of the triangle inequality is applied to the sub template R ^q _{(i, j)} , the template partial region T ^q _{(k, l)} , and S ^q _{(i, j)} (u, v), the following equation (6) Holds.

  Based on the above principle, the lower limit value bound (x, y) expressed by the following equation (7) with respect to (x, y) satisfying u + jm <x <= u + i and v + jn <y <= v + j ) Is updated.

The results of evaluating how fast the template matching apparatus according to the above-described embodiment of the present invention is compared with the conventional method will be described below.
In this example, the image shown in FIG. 4 is an image having 8 bits for each of RGB and having a CIF size (352 × 288) pixels, and five kinds of objects (people, cats, frogs, daruma, bears) fitted inside the image. ) Was used as a template. Each template has a size of 50 × 50 pixels. As the input image, ten artificial images with an S / N ratio of 20 dB generated by adding Gaussian noise to the image shown in FIG. 4 were used.
In the experiment, an Intel Pentium4 (registered trademark) processor (3.06GHZ) was used as the CPU, a Red Hat Linux 8.0 computer as the OS, GNU g ++ as the compiler, and -03 as the optimization option. Used file. Pentium 4 SSE2 instructions were used for important parts of the code.

  According to the embodiment of the present invention, the search time varies depending on the size of the sub-template to be selected. Consider a sub-template that is (m, n) smaller than the template size. As (m, n) is larger, the range in which collation can be omitted increases. On the other hand, since the similarity of the template is smaller <the distance is larger, the lower limit value is smaller and the possibility that collation can be omitted is reduced. In order to examine this trade-off, the size of the sub-template was changed as (m = n = 0,1,2, ..., 9) and the search time was examined. However, (m = n = 0) is consistent with the SSDA method. The sub template for calculating the partial distance value was the upper left sub template R (o ,.), and the search was performed in order from the lower left to the upper right.

  The experimental result by the above is shown in FIG. As a result, for any object, the search time was minimized when a subn plate as small as 3 to 7 was used. Even when sub-templates smaller than 5 were used, the change in search time was small. For this reason, it is considered that a margin size of 3 to 4 should be selected.

In addition, the size of the sub template was set to the farthest right, the initial threshold value was set to 150,000, and the ratio was changed from 10 ⁻³ to 100 ⁰ to measure the search time. The experimental results are shown in FIG. As a result, the tendency of approximation increases as the threshold ratio decreases, and the search time for both the AEJO method and the proposed method is shortened. On the other hand, according to the embodiment of the present invention, detection omission occurred when it was smaller than 10 ⁻² .
This is in the embodiment of the present invention as compared to the case where the AEJO method does not occur other than the detection omission by the sub template selected by the sub template selection unit 9. This is because a detection omission due to another sub-template occurs. The method according to the embodiment of the present invention was always faster than the AEJO method in the range where the ratio of the initial threshold value at which the detection rate becomes 100% is larger than 10 ⁻² .

  As described above, according to the embodiment of the present invention, it can be ensured that the same result as that of the prior art is obtained, and further, an excellent effect that the target signal can be detected at a higher speed than the prior art can be obtained. As a result, when the present invention is applied to an image, it is useful for searching for contents such as a person and an object, tracking a moving object from a moving image, and detecting corresponding points in stereo matching. Further, when the present invention is applied to sound / video information, it is useful for increasing the accuracy and speed of searching for a CM or a desired scene.

  FIG. 10 is a block diagram showing another embodiment of the present invention. Here, the template matching apparatus of the present invention includes a multi-template input unit 101, a sub-template generation unit 102, a reference sub-template determination unit 103, a multi-template grouping unit 104, a self-distance value calculation unit 105, an input Signal capturing unit 106, local region setting unit 107, local region order setting unit 108, reference sub template selection unit 109, candidate local region selection unit 110, distance lower limit value evaluation unit 111, partial distance value calculation unit 112, an additional distance value calculation unit 113, a detection result update unit 114, and a local region pruning unit 115, which input a large number of templates and an input signal, and a signal similar to the template is included in the input signal. And if it is included, the position and distance value are output.

  In FIG. 10, the solid line arrows indicate the flow of processing, the dotted line arrows indicate the data flow, and each functional block is configured by software or hardware, or by a combination of software and hardware. It can be realized. FIG. 11 shows the processing procedure, and FIG. 12 shows an example of a template image. FIG. 13 shows examples of templates processed in each of the multi-template input unit 101, the sub-template generation unit 102, and the reference sub-template determination unit 103, and FIGS. An example of a template processed in each of the conversion unit 104 and the self-distance value calculation unit 105 is shown in (a) to (c).

Here, as shown in FIG. 12 as an example of a template, a target three-dimensional object is searched from an input image using a plurality of templates. Here, the template and a part of the area in the input signal are evaluated by the distance value. First, a method for finding the most similar local region by giving an appropriate threshold value and updating the minimum value d _min found during the search as a threshold value will be described. Of course, it is easy to configure not to update the threshold value, but to record all the values below the threshold value and return the result. In addition, although a two-dimensional example is used here, it is clear that a one-dimensional signal can be handled in the same way by considering a one-dimensional input signal composed of F elements as a two-dimensional 1 × F. It is possible to handle signals of more than three dimensions with the same method. In addition to images, the present invention can also be applied to multidimensional vector sequences such as weather data, map data, and gene data.

  Details will be described below. The multi-template input unit 101 is a two-dimensional image file given in various image formats or a part of one frame image cut out from a moving image file given in various formats, one frame captured in real time. An object region is extracted from a part of the image using a window function or the like, and sent to the sub-template generation unit 102 (S111). For example, the window function of an object can be raised such as a rectangular shape that takes out only the inside of it, a circular shape that takes out only the inside of a circle, or a shape that matches the shape of the object and takes out only the inside of the circle. . Here, an example of a rectangular template will be described as shown in FIG. Further, it is assumed that the template is systematically created by the following method. That is, the object is rotated around a certain axis, the scale is changed stepwise, and so on. The former can be obtained with a computer or the like by placing an object on a rotating sound whose rotation angle can be controlled and photographing the object while rotating the turntable. In the latter case, a desired template can be obtained by changing the zoom parameter of the zoom camera or by changing the distance between the camera and the object and photographing the holiday. It can also be obtained by enlarging or reducing the obtained image by computer graphics processing.

The sub-template generation unit 102 generates an internal sub-template T _sub from all the templates T obtained by the multi-template input unit 1 (S112). As shown in FIG. 13B, suppose a sub-template smaller by (m, n) than the template size. The sub template at position (k, l ⁾ is expressed as T _sub ^{(k, j)} with the upper left corner of the template as the origin. At this time, as shown in FIG. 15, a large number of sub-templates T _sub ^{(k, j)} ... T _sub ^{(m−l, n−l)} that allow overlapping are obtained. The total number is mxn per template. All these sub-templates are sent to the reference sub-template determination unit 103.

The reference sub template determination unit 103 determines a reference sub template T _sub ^{(i, j)} for each template from the multiple sub templates generated by the sub template generation unit 102 (FIG. 13 (c) and S113). At this time, the sub-template at the upper left corner or the lower right corner is used. Further, when searching for a single template from a large number of input images, the maximum value of the distance values between all the sub-templates in the group in the self-distance value calculation unit 105 is set to A sub template that is minimized can be selected as a reference sub template. Let Q be the set of reference sub-templates generated here.

The multi-template grouping unit 104 groups a large number of templates systematically generated by the multi-template input unit 101 according to the similarity of angles or scales, and sends them to the self-distance value calculation unit 105 (S114). In the embodiment of the present invention, for each template, the maximum number of templates in a group in which templates T ′ similar to the template T are assembled is N. Here, there are as many groups as there are multiple templates (FIG. 14A).
The self-distance calculation unit 105 calculates the distance between the reference sub template received from the reference sub template determination unit 104 and other sub templates in the group (FIG. 14B and S115). Here, an absolute value distance, a square distance, or the like can be used as a distance scale, and the above-described normalized cross-correlation value can also be used as a similarity scale. In the following example, an absolute value distance represented by the following formula (8) is used.

Distances d (T _sub ^{(i, j)} , T _sub ^{(k )} between the reference sub template T _sub ^{(i, j)} received from the reference sub template unit 104 and all the sub templates T _sub ^{(k, l)} in the group. ^{, l)} ) is calculated across the template. Further the distance ^{_{d (T sub (i, j}} ), T sub (k, l)) and the distance ^{_{d (T sub (i, j}} ), T sub (k, l)) local region pruning the maximum value of The data is sent to the unit 115 (S115).
The input signal capturing unit 106 is a part of a one-frame image cut out from a two-dimensional image file given in any of various image formats or a moving image file given in any of various formats, or All or part of the one-frame image captured in real time is accepted, and the obtained pixel value vector W is sent to the local region setting unit 107 (S116). The local region setting unit 107 creates a set of all local regions to be collated in the input image received from the input signal capturing unit 106, and sends them to the local region order setting unit 108 (S117). The local area is the same size as the template. A local region having the same size as the template located at the coordinates (u, v) on the input image W is denoted as W ^{(u, v)} . Here, when the set of local regions is W = {S (u, v) | u, v}, the distance lower limit value bound (u, v) is set for all W ^{(u, v) in} W. Then, a position W ^{(u, v)} where the distance d (T, W ^{(u, v)} ) is minimized is found. The initial value of bound (u, v) for all (u, v) is set as appropriate in the implementation, but will be described as bound (u, v) = 0 as an example below.

The local region order setting unit 108 sets the order of local regions (the order in which the input image is scanned) matched with W (S118). For example, when the reference sub template generation unit 102 extracts and collates the upper left region T _sub ^{(i, j)} of the template as a reference sub template, scanning is performed in the order from “lower right to upper left”. Conversely, when collation is performed using the lower right region T _sub ^{(i, j) _ (k, l)} of the template as a reference sub template, scanning is performed in the order of “upper left to lower right”. In the case of the upper left template, it is possible to omit the local region shifted in the upper left direction, and in the case of the lower right sub template, it is possible to omit the local region shifted in the lower right direction. Because it does. On the contrary, the effect of omitting the search is not seen when scanning.

  On the other hand, for speeding up, when collating using a reference sub-template near the center, it is more effective to sequentially select local regions that have been rearranged so as to have an appropriate interval. This method includes a method of rearranging as needed so as to have an appropriate interval based on the distribution of distance values, and a method in which the positions of the centers of pyramid images obtained by recursively dividing the input signal are arranged in order from the largest image. The method used for the local region candidate order and the continuous local region candidate order (upper left to lower right, etc.) are expressed in binary numbers, and the values lower than the most significant 1 are bit-inverted and rearranged in ascending order. There are methods.

The reference sub template selection unit 109 extracts the reference sub template T _sub ^{(i, j)} in order from the reference sub template set Q determined by the reference sub template determination unit 103, and sends it to the candidate local region selection unit 110 (S119). ). When there are no more reference subtemplates included in the set Q, the search is terminated.
When the candidate local region selection unit 110 receives the reference sub template T _sub ^{(i, j)} from the reference sub template selection unit 109, the candidate local region selection unit 110 may include a signal similar to the template T from the local region set W. The region W ^{(u, v)} is selected in the order set by the local region order setting unit 109 and sent to the distance lower limit value evaluation unit 111. When there is no more local area included in W, the selection process by the reference sub template selection unit 109 is executed again.

The distance lower limit evaluation unit 111 compares the distance lower limit value bound (u, v) with the threshold value d _min for W ^{(u, v)} extracted by the candidate local region selection unit 110 (S121). At this time, if bound (u, v)> d _min , that is, if the distance lower limit value is larger than the threshold value, the candidate local region selection unit 110 performs candidate local region selection again. If the comparison results in bound (u, v) ≦ d _min , the distance lower limit evaluation unit 111 may include a signal similar to the template T in the candidate local region W ^{(u, v).} Therefore, the position (u, v) is sent to the partial distance value calculation unit 112.

The partial distance value calculation unit 112 has the same position and size as the reference sub template T _sub ^{(i, j)} received from the reference sub template selection unit 109 and W ^{(u, v)} extracted by the candidate local region selection unit 110. A distance d (T _sub ^{(i, j)} , W _sub ^{(u, v)} ) is calculated with respect to a partial area W _sub ^{(u, v)} (here, referred to as a partial local area) (S122). .
In this distance calculation process, the collection of images for which distance value calculation has been completed monotonously increases, so the distance armor also monotonously increases. Therefore, the partial distance value calculation unit 112 sequentially evaluates whether or not the distance value exceeds the threshold value during the calculation, and determines whether or not to terminate the calculation. By configuring in this way, the SSDA method can be used in which unnecessary distance calculation is omitted. Here, d _min + d _max is used as the threshold value. When the distance value exceeds d _min + d _max , the calculation is terminated because the target object is not included in the partial candidate local area W _sub ^{(u, v)} and the surrounding partial candidate local areas. Further, when the distance value exceeds d _min after all the distance calculations are completed, similarly, the target local area W ^{(u, v)} does not include the target object.

When the distance value is greater than or equal to d _min , the partial distance value calculation unit 112 sends the distance value d to the local region pruning unit 115 in order to check whether an object is included in the W ^{(u, v)} peripheral region. Send it out.
On the other hand, after all the distance calculations are completed, if the distance value is equal to or less than d _min , there is a possibility that the object is the target object. To send. The additional distance value calculating unit 113 adds the distance between the remaining part of the template and the remaining part of the local region to the distance value of the region calculated by the partial distance value calculating unit 112, thereby Find the distance between.

Here, if the region on the template is T and the region not included in the reference sub template T _sub ^{(i, j)} is ~ T _sub ^{(i, j)} , the region obtained by subtracting the region of the reference sub template from the template Is T∩ ~ T _sub ^{(i, j)} . Similarly, the area obtained by subtracting the partial local area W _sub ^{(u, v)} from the local area W ^{(u, v)} is W ^{(u, v)} ∩ to W _sub ^{(u, v)} .
That is, the partial distance value d (T _sub ^{(i, j)} , W _sub ^{(u, v)} ) obtained by the partial distance calculation unit 112 is added to the distance d (T∩˜T _sub ^{(i, j} ) between them. ⁾ , W ^{(u, v)} ∩ ~ W _sub ^{(u, v)} is added to obtain the distance d (T, W ^{(u, v)} ) between the template and the candidate local region. The distance d (T, W ^{(u, v)} ) of the local region is expressed by the following equation (9).

At this time, the SSDA method is used similarly to the partial distance value calculation unit 112. Then, as a result of sequentially calculating the distance value, if it exceeds d _min in the middle, it is not the target object, so the additional distance value calculation unit 113 interrupts the distance value calculation, and the candidate local region selection unit 110 The next candidate local region is selected at. On the other hand, when calculating to the last pixel and the distance between the template and the distance d between the local areas does not exceed d _min , the additional distance value calculation unit 113 sends the distance d to the detection result update unit 114.
The detection result update unit 114 updates the minimum value d _min with the distance value d obtained from the additional distance value calculation unit 113 (d _min = d), and records the position (u, v) (S124). If the set W is not an empty set, candidate local region selection processing by the candidate local region selection unit 110 is executed, and the next candidate local region is selected.

The local region pruning unit 115 performs an update process of the distance lower limit value bound (u, v) between the template and the surrounding local region from the distance relationship between the reference sub template and a part of the local region. In the local region where a large distance lower limit value bound (u, v) is obtained, the distance lower limit value evaluation unit 111 checks in the local region where bound (u, v)> d _min is satisfied. Can be omitted. Therefore, the search can be speeded up without sacrificing accuracy.

When the distance between the reference sub template and the partial region of the local region is calculated based on the self-distance value of the sub-template of the template similar to the reference template obtained by the self-distance value calculation unit 105, it extends over other templates. Thus, the lower limit of distance can be calculated. The principle is shown below.
The distance value to which the SSDA method is applied is obtained by sequentially adding the distance for each pixel of the local region and each pixel of the sub template. That is, as the number of pixels to be compared increases, the distance value also increases monotonously. For this reason, the distance d (T _sub ^{( )} between the reference sub template T _sub ^{(i, j)} as shown in FIG. 15 and a part W _sub ^{(u, v)} of the local region having the same size as the reference sub template. ^{i, j)} , _Wsub ^{(u, v)} ) and the distance d (T, W ^{(u, v)} ) between the template and the local region, the following equation (10) holds.

In other words, if the reference sub-template does not match as a result of matching, the entire template is not matched. Further, applying the principle of triangular inequality with respect to the reference sub template T _sub ^{(i, j)} , other sub templates T ′ _sub ^{(i, j)} in the same group, and W _sub ^{(u, v)} , the following equation The relationship shown by (11) is established.

Using the matching result of the reference sub template T _sub ^{(i, j)} and the partial candidate local region W _sub ^{(u, v)} , the lower limit value of distance when matching with another template T ′ is calculated in the same way. Can do. For example, it is represented by the following formula (12).

  Based on the above principle, the lower limit value bound (x, y) is updated for (x, y) that satisfies u−m <x <= u and v−n <y <= v. The update can be expressed by the following equation (13).

The results of evaluating how fast the above-described embodiment of the present invention is compared with the conventional method will be described below.
Here, the search time when the margin m = n and the number N of templates to be processed simultaneously was changed was measured and compared with other methods. The experiment was divided into two types: a template in which only the rotation angle was changed, and a template in which only the scale was changed. In the experiment, an Intel XEON CPU (2.2 GHz) as a CPU, a 3 GB memory, a Microsoft Windows XP (registered trademark) Professional (SP1) computer as the OS, and MATLAB as the software were used.

In the former “search using a template with a changed rotation angle”, 36 templates obtained by photographing the target object every 10 degrees were used. The template size was 48 × 48 pixels, and the input image size was 320 × 320 pixels. The threshold is a value 1 × 10 ⁵ obtained by performing a search from 360 input images photographed every other time using 324 images having different angles from the template and measuring the threshold value where Precision = Recall = 100%. Set.
The search time (the object is “fuku” shown in FIG. 17 (d)) when the present invention is applied while changing N and m described above is graphically displayed in FIG.
With respect to the margin, it can be seen that the search time is the shortest when m is around 13, and the number of templates to be processed simultaneously decreases rapidly until N reaches around 3, and there is almost no change thereafter. FIG. 19 shows the search time when m is changed at N = 11 where the search time is the shortest. According to FIG. 19, as m increases, the search time decreases rapidly, becomes the shortest around 9 [pixel] to 13 [pixel], and then increases. If the margin is increased, the area that can be omitted by one verification increases, but the distance between sub-references increases and the possibility of the verification omitted decreases. The search time is reduced by increasing m for simple objects.

Similarly, FIG. 20 shows the search time when N is changed at m = 13 where the search time is the shortest in the example shown in FIG.
According to FIG. 20, the search time decreases rapidly to near N = 4 and thereafter changes little. Although each object shows the same decreasing tendency, it can be confirmed that the N including the shortest search time decreases as the image includes a high-frequency component.

  For comparison, the search time by STM, SSDA, subspace method, and template matching method was measured. STM and SSDA used the above threshold values. In the subspace method, the search time was measured with the minimum number of dimensions such that Precision = Recall = 100%. The result is shown as <Table 1> in FIG. Both are times when the search time is the shortest. According to the method of the present invention, it can be seen that the search is completed at a speed approximately 2.7 times that of the STM method and 7 times that of the partial empty method.

  The latter “search using a template with a changed scale” will be described. Here, a 48 × 83 pixel image cut out from the photographed image and reduced to (15/16) k was used as a template. However, k = 1, ..., 14. The input image is a 320 × 240 photographed image reduced to 70% vertically and horizontally. Further, the threshold value was set to 30 per pixel by measuring the threshold value at which Precision = Recall = 100%.

The object is “fuku”, and the result is displayed in a graph in FIG. According to FIG. 21, with respect to the margin, the search time is the shortest when m is about 9, and the number of templates to be processed simultaneously decreases rapidly until N reaches around 3, and then increases gradually. I understand that. FIG. 22 shows the search time for each object when m is changed at N = 3 where the search time is the shortest from FIG. From FIG. 22, it can be seen that when m is from 7 to 9, the search time is the shortest for each object.
Similarly, FIG. 23 shows the search time for each object when N is changed in the vicinity of m = 9 where the search time is the shortest. According to FIG. 23, the search time is shortest around N = 3 for each object, and then warmly increases.

Further, for comparison with other methods, was measured STM, SSDA, the search time by the template matcher switch ing method. Here, the above-described threshold values are used for STM and SSDA. The result is shown as <Table 2> in FIG. Both are times when the search time is the shortest. According to <Table 2>, it can be seen that the search is completed at a speed approximately twice that of the STM method and 20 times that of the SSDA method.

As described above, according to the embodiment of the present invention, the template collation method can be extended so that a three-dimensional object can be searched at high speed using a large number of templates whose rotation angles and scales are changed. Specifically, a plurality of templates can be searched together, and templates of different sizes can be handled together. As a result of experiments on a rotating object, a speed increase of about 2.7 times that of the conventional STM method and about 7 times that of the partial empty method can be realized. Further, it is possible to realize a speed increase of about twice as high as that of the STM method and about 20 times as high as that of the SSDA method for an object whose photographing size changes.
As a result, when the present invention is applied to an image, it is useful for searching for contents such as a person and an object, tracking a moving object from a moving image, and detecting corresponding points in stereo matching. In addition, when applied to sound / video information, it is useful for increasing the accuracy and speed of searching for a CM or a desired scene.

It is a block block diagram which shows one Embodiment of this invention. It is the flowchart quoted in order to demonstrate operation | movement of one Embodiment of this invention. It is an example of the template concerning this embodiment. It is an example of the input signal concerning this embodiment. It is a figure which shows an example of the result of having detected the area | region similar to a template from an input signal. It is a figure which shows an example of the stepwise sub template contained in a template. It is the operation | movement conceptual diagram quoted in order to demonstrate operation | movement of one Embodiment of this invention. It is the figure which displayed the experimental result which showed the change of the search time when changing the size of a sub template in the graph. It is a figure showing the experimental result which showed the change of the search time of the conventional method and this invention when a threshold value was changed. It is a block block diagram which shows other embodiment of this invention. It is the flowchart quoted in order to demonstrate operation | movement of other embodiment of this invention. It is a figure which shows an example of the object searched in other embodiment of this invention. It is the figure quoted in order to demonstrate the relationship between a part of structure shown in FIG. 10, and a template. It is the figure quoted in order to demonstrate the relationship between a part of structure shown in FIG. 10, and a template. It is a figure which shows an example of the sub template contained in a template. It is the operation | movement conceptual diagram quoted in order to demonstrate the operation | movement in other embodiment of this invention. It is a figure which shows an example of the three-dimensional object searched in other embodiment of this invention. It is the graph quoted in order to demonstrate the search time of the rotating object (fuku) shown in FIG. It is the figure which displayed the relationship between the margin dependence of search time, and a rotating object in the graph. It is the figure which displayed the relationship between the group size (N) dependence of search time, and a rotating object on the graph. It is the graph quoted in order to demonstrate the search time of the size change object (fuku) shown in FIG. It is the figure which displayed the relationship between the margin size dependence of search time and a size change object in the graph. It is the figure which displayed the relationship between the group size (N) dependence of search time and a size change object by the graph display. It is the figure which showed contrast with embodiment of this invention and a prior art example by tabular form regarding search time.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Template input part, 2 ... Sub template production | generation part, 3 ... Gradual sub template production | generation part, 4 ... Self distance value calculation part, 5 ... Input signal capture part, 6 ... Local area | region setting part, 7 ... Local area | region order setting , 8 ... Candidate local area selection section, 9 ... Sub template selection section, 10 ... Partial distance value calculation section, 11 ... Additional distance value calculation section, 12 ... Detection result update section, 13 ... Candidate local area pruning section, 101 DESCRIPTION OF SYMBOLS ... Multi template input part, 102 ... Sub template production | generation part, 103 ... Reference sub template determination part, 104 ... Multi template grouping part, 105 ... Self distance value calculation part, 106 ... Input signal capture part, 107 ... Local area setting part , 108 ... Local region order setting unit, 109 ... Reference sub-template selection unit, 110 ... Candidate local region selection unit, 111 ... Distance lower limit value Valence unit, 112 ... partial distance value calculation unit, 113 ... Add distance value calculation unit, 114 ... detection result update unit, 115 ... candidate local region pruning unit

Claims (11)

A template matching device that detects a signal similar to a known signal given as a template from an input signal that is input and captured, and obtains its position,
A sub-template generation unit that generates sub-templates of different sizes from the template in stages;
Calculating a distance value between the input signal and the sub-template sequentially generated by the sub-template generation unit, and when it is confirmed that the similarity is similar, a matching calculation unit that additionally calculates a distance value between the entire template;
Have
The matching operation unit
A self-distance value calculation unit for calculating a distance value with respect to all the sub-templates having different sizes in steps, with respect to any partial region having the same size as the sub-template on the template;
On the input signal, a local region setting unit that sets any local region having the same size as the template,
A local region order setting unit for rearranging the order of the local regions;
One local region is extracted in the order set by the local region order setting unit as a candidate local region, and the lower limit distance between the candidate local region and the template is compared with a predefined threshold value. A candidate local region selection unit for selecting a candidate local region;
A sub template selection unit that selects one sub template in ascending order of size from a plurality of sub templates obtained by the stepwise sub template generation unit;
A partial distance value calculation unit that calculates a partial distance value by sequentially comparing a partial region of the candidate local region obtained from the candidate local region selection unit and a sub template obtained by the sub template selection unit;
If the partial distance values obtained by the partial distance value calculation unit are similar, add the distance between the remaining area of the template and the remaining area of the candidate local area, and between the template and the candidate local area An additional distance value calculation unit for obtaining a distance value between and
When the distance value obtained by the additional distance value calculation unit is below a target threshold value, the detection result and a detection result update unit that updates the threshold value;
Similar to the template by updating the distance lower limit value of the local region in the vicinity of the candidate local region, using the respective distance values calculated and output by the self-distance value calculating unit and the partial distance value calculating unit. A candidate local region pruning unit that deletes local region candidates that may not have a signal to be
The template collation apparatus characterized by comprising. A template matching device that detects a signal similar to a known signal given as a template from an input signal that is input and captured, and obtains its position,
A sub-template generation unit that generates a plurality of sub-templates of a predetermined size from the template;
A reference sub-template determination unit that determines, for each template, a reference sub-template for matching from the sub-template,
A grouping of templates having similar characteristics among the templates, a self-distance value of a sub-template straddling the template in the group is calculated in advance, and a result of matching with the certain reference sub-template is reflected in the group Calculation unit and
Have
The matching operation unit
A self-distance value calculating unit that calculates all the distance values between the reference sub-template determined by the reference sub-template determining unit and the sub-templates in the same group across the template;
A local area setting unit for setting any local area of the same size as the template on the input signal;
A local region order setting unit for rearranging the order of the local regions;
A reference subtemplate selection unit that sequentially extracts reference subtemplates from the set of reference subtemplates determined by the reference subtemplate determination unit;
A candidate local region selection unit that takes out one local region in the order set by the local region order setting unit and sets it as a candidate local region;
A distance lower limit evaluation unit that compares a distance lower limit between the candidate local region selected by the candidate local region selection unit and the template with a predefined threshold;
A partial distance value calculation unit that compares a part of the candidate local region obtained from the candidate local region selection unit with the reference subtemplate selected by the reference subtemplate selection unit, and calculates a partial distance value;
If the partial distance values obtained by the partial distance value calculation unit are similar, add the distance between the remaining area of the template and the remaining area of the candidate local area, and between the template and the candidate local area An additional distance value calculator for calculating the distance value of
A detection result updating unit for updating the detection result and the threshold when the distance value obtained by the additional distance value calculation unit is lower than a target threshold;
Using the distance values calculated by the self-distance value calculation unit and the partial distance value calculation unit, there is a signal similar to the template by updating the distance lower limit value of the local region near the candidate local region. A candidate local region pruning unit that deletes local region candidates that are not likely to
The template collation apparatus characterized by comprising. The partial distance value calculator and the additional distance value calculator are
Said distance value sequentially calculated and approximately no computational need of, or already template matching apparatus according to claim 1 or 2, characterized in that the calculation ends when the need is found not in the calculation . The sub-template generation unit
The maximum value of the self-distance value calculated by the self-distance calculation section, template matching apparatus according to claim 1, wherein the generating stepwise sub-template is minimized. The reference sub- template determination unit
3. The template matching apparatus according to claim 2 , wherein a sub template having a minimum maximum distance value between all sub templates in the group is determined as a reference sub template. The local region order setting unit includes:
Template matching apparatus according to claim 1 or 2, characterized in that at predetermined intervals rearranges the order. The local region order setting unit includes:
Template matching apparatus according to claim 1 or 2, characterized in that on the basis of the distribution of the distance values, rearranged to have a suitable distance from time to time. The local region order setting unit includes:
The template collation apparatus according to claim 1 or 2 , wherein the center positions of the pyramid images obtained by recursively dividing the input signal are arranged in order from a large image, and used in the order of extracting the candidate local regions. The local region order setting unit includes:
The local region candidate order is expressed in binary notation, and the value lower than the highest one is bit-inverted, arranged in ascending order or descending order, and used in the order of extracting the candidate local areas. The template collation apparatus according to claim 1 or 2 . A template collation method for operating a template collation apparatus that detects a signal similar to a known signal given as a template from an input signal that is inputted and captured and determines its position,
A first step in which a sub-template generation unit generates sub-templates having different sizes from the template in stages;
A matching calculation unit calculates a distance value between the input signal and the sequentially generated sub-template, and when it is confirmed that the similarity is similar, a second step of additionally calculating a distance value with the entire template;
Have
The second step includes
A sub-step in which a self-distance value calculation unit calculates, for all sub-templates having different sizes in stages, a distance value between all sub-regions having the same size as the sub-template on the template;
A sub-step in which a local region setting unit sets any local region having the same size as the template on the input signal;
A local region order setting unit, a sub-step of rearranging the order of the local regions;
The candidate local region selection unit extracts one local region in the set order as a candidate local region, compares the lower limit of the distance between the candidate local region and the template with a predefined threshold value, and then compares them. A sub-step of selecting candidate local regions to be
A sub-step for selecting one sub-template from the plurality of sub-templates in ascending order of size;
A partial step of calculating a partial distance value by sequentially comparing a partial area of the selected candidate local area and the selected subtemplate;
If the obtained partial distance value is similar, the additional distance value calculation unit adds the distance between the remaining area of the template and the remaining area of the candidate local area, and the template and the candidate local area A sub-step for obtaining a distance value between
When the detection result update unit is below the target threshold value, the sub-step of updating the detection result and the threshold value,
The candidate local region pruning unit updates the distance lower limit value of the local region in the vicinity of the candidate local region using the calculated and output self-distance value and partial distance value, so that a signal similar to the template is obtained. A sub-step of removing local region candidates that may not exist;
A template matching method characterized by comprising: A template collation method for operating a template collation apparatus that detects a signal similar to a known signal given as a template from an input signal that is inputted and captured and determines its position,
A first step of generating a plurality of sub-templates of a predetermined size from the template;
A second step in which a reference subtemplate determination unit determines a reference subtemplate to be matched from the subtemplate for each template;
The matching calculation unit groups templates having similar characteristics among the templates, calculates in advance the self-distance value of the sub-template straddling the template in the group, and compares the result of the reference sub-template with the group. A third step to reflect in,
Have
The third step includes
A sub-step in which a self-distance value calculation unit calculates all self-distance values between the determined reference sub-template and sub-templates in the same group across the template;
A sub-step in which a local region setting unit sets any local region having the same size as the template on the input signal;
A local region order setting unit, a sub-step of rearranging the order of the local regions;
A sub-step in which a reference sub-template selection unit sequentially selects a reference sub-template from the determined set of reference sub-templates;
A candidate local region selection unit takes out one local region in the order set by the local region order setting unit and selects it as a candidate local region; and
A sub-step in which a distance lower limit evaluation unit compares a distance lower limit between the selected candidate local region and the template with a predefined threshold;
A partial step of calculating a partial distance value by collating a part of the selected candidate local region with the selected reference sub-template, and
If the partial distance value is similar, the additional distance value calculation unit adds the distance between the remaining area of the template and the remaining area of the candidate local area, and the distance between the template and the candidate local area A substep of calculating a value;
A detection result updating unit, when a distance value between the obtained template and the candidate local region is below a target threshold value, a sub-step of updating the detection result and the threshold value;
The candidate local area pruning unit uses the calculated self-distance value and partial distance value to update the distance lower limit value of the local area in the vicinity of the candidate local area, so that a signal similar to the template exists. A sub-step of removing local area candidates that are not possible;
A template matching method characterized by comprising:

Images