JP2013069125A - Image retrieval device, image retrieval method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable a user to obtain a retrieval result which reduces mental load of the user by simple processing.SOLUTION: An image retrieval device which calculates a rotation degree in a process of calculating a local feature amount necessary for a retrieval, determines, when displaying an image as a result of the retrieval, the necessity of rotation normalization according to the estimated rotation degree, and displays, when the rotation degree is equal to or more than a predetermined value, a sample image rotationally normalized with reference to a query image. Accordingly, a user can reduce load caused by a mental rotation when confirming the retrieval result, which is obtained by suppressing the amount of calculation.

Description

  The present invention relates to a technique suitable for use for performing a search based on local feature points and local feature amounts, for example.

  Conventionally, a method of searching for a similar image using a local feature amount (local feature amount) of an image has been proposed. In this method, first, characteristic points (local feature points) are extracted from an image. Examples of the extraction method include the method disclosed in Non-Patent Document 1. Then, based on the feature point and the surrounding image information, a feature amount (local feature amount) corresponding to the feature point is calculated. The local feature amount can be calculated by a method disclosed in Non-Patent Document 2, for example. In this way, the image search is performed by matching local feature amounts.

  In the method using the local feature amount, the local feature amount is defined as information including a plurality of elements that are rotation invariant and enlargement / reduction invariant. Thereby, even when the image is rotated, enlarged or reduced, the image can be searched. Further, the local feature amount is generally expressed as a vector. However, it is a theoretical story that the local feature is invariant to rotation and enlargement / reduction. In an actual digital image, the local feature before image rotation and enlargement / reduction processing corresponds to that after processing. Some variation occurs between the local feature amount.

  For example, in the method described in Non-Patent Document 2, in order to calculate a rotation-invariant local feature amount, a main direction is calculated from a pixel pattern of a local region around a local feature point, and the main direction is used as a reference when calculating the local feature amount. Rotate the local region to normalize the direction. In addition, in order to calculate the local feature amount that does not change in size, the image is generated internally with different resolutions, local feature points are extracted from each resolution image, and the local feature amount is calculated. Here, a series of images having different resolutions generated inside is generally called a scale space.

  Further, in the conventional search result display method, the similarity between the query image and the search destination image (sample image) is calculated by matching local feature amounts, and the thumbnail images are displayed in the order of similarity.

  On the other hand, when the search result is displayed in this way and it is perceived that the rotating objects are the same, there is an established theory that matching is performed after creating a rotated image in the head. Hereinafter, such a feeling is called mental rotation. Non-Patent Document 3 describes that the perception time accompanied by mental rotation is proportional to the rotation angle of the target object, and in the case of 180 ° rotation, it takes about twice as long as in the case of no rotation. Has been.

  When confirming a large number of images displayed as search results, this mental rotation may increase the perception time and increase the burden on the user. Therefore, an example of rotating and displaying a rectangle surrounding an object is known. However, when the perceived object (search result) is rotating, it can be said that the processing in consideration of the burden on the user is simply surrounded by the rectangle. Absent.

  As another example of rotating the image, after determining the top / bottom direction of the image including the person image and correcting the image to the actual top / bottom direction based on the determination result, there is a human image in the image. A method for extracting a region to be performed has been proposed (see, for example, Patent Document 1).

  Furthermore, a method for correcting the inclination of a moving image frame has also been proposed (see, for example, Patent Document 2). In this method, first, using a motion vector, a region constituting a frame image is divided into a region having a relatively large amount of motion and a region having a relatively small amount of motion. Next, a correction parameter is detected from a region where the amount of motion is relatively small. Finally, motion correction is performed on the frame image. In this method, the inclination determination process is always performed regardless of the degree of inclination.

JP 2005-215760 A International Publication No. 2009/001510

C. Harris and M.J. Stephens, "A combined corner and edge detector," In Alvey Vision Conference, pages 147-152, 1988. David G. Lowe, "Distinctive Image Features from Scale-Invariant Keypoints," International Journal of Computer Vision, 60, 2 (2004), pp.91-110. Jacques Vauclair, Joel Fagot, William D Hopkins, "Rotation of mental images in baboons when the visual input is directed to the left cerebral hemisphere," Psychological Science, Vol. 4 (2), pp.99-103, 1993. J.J.Koenderink and A.J.van Doorn, "Representation of local geometry in the visual system," Riological Cybernetics, vol.55, pp.367-375, 1987. M. A. Fischler and R. C. Bollers, "Random sample consensus: A paradigm formodel fitting with applications to image analysis and automated cartography," Commum. ACM, no.24, vol.6, pp.381-395, June 1981.

  However, in the method described in Patent Document 1, the top-and-bottom direction of an image including a person is determined, and the image is rotated in units of 90 degrees using the determination result. For this reason, the load caused by mental rotation cannot be reduced depending on the rotation angle of the image.

  Further, in the method described in Patent Document 2, since the tilt determination process is configured as an independent process, the calculation amount increases accordingly. In addition, since the inclination determination process is always performed regardless of the degree of inclination, the processing load increases in accordance with the calculation resource of the mounted device.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to make it possible to obtain search results that reduce the mental load on the user by simple processing.

  The image search apparatus of the present invention collates local feature extraction means for extracting local local features from an input image and a comparison image, and local features of the input image and comparison image extracted by the local feature extraction means. Collation means for calculating similarity, rotation angle calculation means for calculating a rotation angle of the comparison image with respect to the input image based on a value obtained when the similarity is calculated by the collation means, and the rotation angle A rotation unit that rotates the comparison image according to the rotation angle calculated by the calculation unit; and an output unit that outputs the comparison image rotated by the rotation unit according to the similarity calculated by the collation unit. It is characterized by.

  According to the present invention, when a search result is confirmed, the load due to mental rotation can be reduced, and the search result can be obtained with a reduced amount of calculation.

It is a block diagram showing an example of composition of an image search device in an embodiment. In an embodiment, it is a flow chart which shows an example of a processing procedure which extracts a local feature point and local feature-value of an image. It is a figure explaining the outline | summary of the production | generation process of the reduction image in step S204 of FIG. 5 is a flowchart illustrating an example of a processing procedure for outputting a search result in the first embodiment. In the 1st-3rd embodiment, it is a flowchart which shows an example of the process sequence which calculates a similarity and a rotation angle. 10 is a flowchart illustrating an example of a processing procedure for outputting a search result in the second embodiment. In 2nd Embodiment, it is a flowchart which shows an example of the process sequence of rotation normalization when the display result is clicked. 14 is a flowchart illustrating an example of a processing procedure for outputting a search result in the fourth embodiment. In a 4th embodiment, it is a flow chart which shows an example of a processing procedure which computes a similarity and a modification parameter.

(First embodiment)
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram illustrating a configuration example of an image search apparatus 100 according to the present embodiment.
In FIG. 1, an image search apparatus 100 includes a query image input unit 102, a comparison target image input unit 105, a first local feature point / local feature amount extraction unit 103, and a second local feature point / local feature amount extraction unit 106. , And a local feature point / local feature quantity comparison unit 107.

  In the image search apparatus 100, when the query image 101 that is an input image is input to the query image input unit 102, the first local feature point / local feature quantity extraction unit 103 extracts a local feature point from the query image 101 as a local feature. And its local features. Similarly, when the comparison target image sequence 104 to be compared is input to the comparison target image input unit 105, the second local feature point / local feature quantity extraction unit 106 outputs a local feature as a local feature from the comparison target image sequence 104. Feature points and their local features are extracted. Details of the process of extracting the local feature points and the local feature amounts will be described later.

  The local feature point / local feature amount comparison unit 107 includes the local feature points / local features extracted by the first local feature point / local feature amount extraction unit 103 and the second local feature point / local feature amount extraction unit 106. Compare feature quantities. Then, the comparison result 108 is output to a display device (not shown) as an image display order. Details of the comparison processing of the local feature points and the local feature amounts will be described later. The operation unit 109 is for receiving input from the user.

  First, processing of the query image input unit 102, the comparison destination image input unit 105, the first local feature point / local feature amount extraction unit 103, and the second local feature point / local feature amount extraction unit 106 in FIG. 1 will be described. . The local feature extraction process from an image includes a step of reading image data, a luminance component extraction step, a reduced image creation step, a local feature point extraction step, and a local feature amount calculation step.

FIG. 2 is a flowchart illustrating an example of a processing procedure for extracting local feature points and local feature amounts of an image in the present embodiment.
First, processing is started in step S201, and in step S202, the query image input unit 102 and the comparison destination image input unit 105 read out the respective image data. In step S203, the first local feature point / local feature amount extraction unit 103 and the second local feature point / local feature amount extraction unit 106 each extract a luminance component from the image data to generate a luminance component image. To do.

  Next, in step S204, each of the first local feature point / local feature quantity extraction unit 103 and the second local feature point / local feature quantity extraction unit 106 uses the luminance component image created in step S203 according to the magnification p. The image is sequentially reduced to generate n reduced images. However, the magnification p and the number n of reduced images are determined in advance.

  FIG. 3 is a diagram for explaining the outline of the reduced image generation process in step S204 of FIG. FIG. 3 shows an example where the magnification p is 2 to the power of − (1/4) and the number n of reduced images is 9. Further, in the example shown in FIG. 3, the magnification p is not the area ratio but the side length ratio.

  In FIG. 3, reference numeral 301 denotes the luminance component image created in step S203. Reference numeral 302 denotes a reduced image obtained by reducing the luminance component image 301 four times in accordance with the magnification p, and corresponds to an image obtained by reducing the luminance component image 301 to ½. Reference numeral 303 denotes a reduced image obtained by reducing the luminance component image 301 eight times according to the magnification p, and corresponds to an image obtained by reducing the luminance component image 301 to ¼. Note that, as a method for reducing an image, a method of simply thinning out pixels, a method using linear interpolation, a method of sampling after applying a low-pass filter, or the like can be used. Any method may be used. In this embodiment, it is assumed that an image is reduced using a reduction method based on linear interpolation.

  Reference numeral 304 denotes a scale number, which is a number assigned according to the resolution in order from the larger reduced image size. In this embodiment, the scale number starts from 1. However, the scale number may start from 0.

  Returning to the description of FIG. 2, next, in step S205, the first local feature point / local feature quantity extraction unit 103 and the second local feature point / local feature quantity extraction unit 106 obtain the n obtained in step S204. Local feature points (local feature points) are extracted from each of the reduced images. The local feature points extracted here are robust local feature points that can be stably extracted from the same place even if image processing such as rotation or reduction is performed on the image. As a method for extracting such local feature points, the Harris operator described in Non-Patent Document 1 is used in the present embodiment.

  Specifically, for each pixel of the image obtained by applying the Harris operator, the pixel value of a total of nine pixels including the pixel of interest and the eight pixels around it is examined. When the pixel value of the pixel of interest is equal to or greater than a predetermined value and has a local maximum (the pixel value of the pixel is the maximum among the nine pixels), the point where the pixel of interest is located is locally Extract as feature points. Note that any feature point extraction method is applicable as long as it is a method capable of extracting a robust local feature point, not limited to the above-described feature point extraction method using the Harris operator.

  Next, in step S206, for each local feature point obtained in step S205, a feature quantity (local feature quantity) defined so as to remain unchanged even when the image is rotated is calculated. As a method for calculating the local feature amount, in this embodiment, a combination of Local Jet and derivatives thereof described in Non-Patent Document 4 is used. The local feature amount calculated by this method can have characteristics that have a relatively high resistance to enlargement / reduction and rotation. Specifically, the local feature amount is calculated by the following equation (1).

  However, the symbols used on the right side of Expression (1) are defined as shown in Expression (2) to Expression (7) below.

Here, G (x, y) on the right side of Expression (2) is a Gaussian function, I (x, y) is a pixel value at image coordinates (x, y), and “*” represents a convolution operation. It is a symbol. Equation (3) is a partial derivative of variable L defined in equation (2) with respect to x, and Equation (4) is a partial derivative of variable L with respect to y. Equation (5) is a partial derivative with respect to y of the variable L _x defined in Equation (3). Equation (6) is a partial derivative of variable L _x defined in equation (3) with respect to x, and equation (7) is a partial derivative of L _y defined in equation (4) with respect to y.

  In addition, the calculation method of a local feature-value is not restricted to the above-mentioned method, The calculation method of another local feature-value is applicable. As another local feature amount calculation method, for example, there is a method described in Non-Patent Document 2. In this method, the gradient direction of the pixel value is calculated for each pixel in the local region around the local feature point, and the histogram is used as the local feature amount.

  Next, the operation of the local feature point / local feature amount comparison unit 107 will be described. There are various methods for collating images based on the comparison of local feature points and local feature amounts. In this embodiment, a method using RANSAC described in Non-Patent Document 5 will be described.

FIG. 4 is a flowchart illustrating an example of a processing procedure for outputting a search result by the local feature point / local feature amount comparison unit 107 in the present embodiment.
First, processing is started in step S401 in FIG. 4, and one sample image to be compared is extracted from the comparison destination image sequence 104 input to the comparison destination image input unit 105 in step S402.

  In step S403, the similarity between the query image 101 input to the query image input unit 102 and the sample image is calculated. Further, a rotation angle of the sample image with respect to the query image is estimated using a value generated in the process of calculating the similarity. Details of the processing in step S403 will be described later. The similarity and rotation angle calculated at this time are linked to the sample image.

  Next, in step S404, it is determined whether or not there is an image that has not yet been extracted as a sample image in the comparison target image sequence 104. As a result of this determination, if an image remains, the process returns to step S402, a new sample image is extracted, and the same processing as described above is performed. On the other hand, if no image remains, the process proceeds to step S405. In step S405, the sample images are sorted in the order of similarity using the similarity of each sample image to the query image calculated in step S403.

  Next, in step S406, it is determined whether rotation normalization of the sample image with respect to the query image is necessary. As a result of the determination, if rotation normalization is necessary, the process proceeds to step S407, and if rotation normalization is not necessary, the process proceeds to step S408. Here, various criteria can be considered as the rotation normalization determination criteria. In this embodiment, it is determined that rotation normalization is required when the estimated rotation angle is ± 60 ° or more. This is based on the fact that the experiment described in Non-Patent Document 3 shows that the increase in processing time due to mental rotation is not so large within the range of the rotation angle of ± 60 °.

  Further, if the determination criterion in step S406 is a reasonable criterion, a criterion other than the criterion of the present embodiment may be adopted. For example, a rotation normalization process may be performed on all of a specific number of images in order from the highest similarity. Further, it may be configured to determine whether rotation normalization is necessary based on a predetermined similarity value, or to determine whether rotation normalization is required by obtaining a distribution of similarity values. The similarity threshold may be relatively determined. Furthermore, the threshold value may be dynamically determined according to the resource of the device.

  Furthermore, a function for determining the user's cognitive ability may be provided, and a threshold value may be determined in advance or dynamically based on the determination result. As a judgment standard of cognitive ability, age, reaction time, visual acuity, etc. are mentioned, for example. As an age discrimination method, for example, an age estimation technique from an existing face image can be used. As a method for measuring the reaction time, for example, a result of a separately prepared reaction time test program may be used, or a method of analyzing user input data from a mouse or keyboard may be used. As for the visual acuity measurement method, for example, a result of a separately prepared visual acuity measurement test program may be used, or the user himself / herself may input a numerical value of visual acuity.

  In step S407, the rotation normalization process based on the query image is performed on the sample image using the rotation angle estimated in step S403. In step S408, the comparison target image sequence sorted in the order of similarity and subjected to the rotation normalization process as necessary is output as the comparison result 108.

Next, details of the processing in step 403 will be described. Here, the local feature point extracted from the query image 101 is Q, its coordinates are Q (x ′, y ′), and the local feature amount of the local feature point Q is V _q . In addition, the local feature point on one image (sample image) belonging to the comparison target image sequence 104 is set as S, the coordinates thereof are set as S (x, y), and the local feature amount of the local feature point S is set as V _s. And

FIG. 5 is a flowchart showing an example of a processing procedure for calculating the similarity and the rotation angle by the local feature point / local feature amount comparison unit 107, which is the detailed processing of step S403.
Upon starting the processing step S501, in step S502, to create a minimum distance correspondence point list by matching the local feature quantity V _q and the local feature amount V _s. Specifically, first, distances between local feature quantities are calculated for all combinations of the local feature quantity V _q and the local feature quantity V _s . Next, a combination (corresponding point) of the local feature quantity V _q and the local feature quantity V _s is extracted so that the calculated distance between the local feature quantities is equal to or less than the threshold value Tv and is the minimum distance. By registering these corresponding points in the minimum distance corresponding point list, a minimum distance corresponding point list is created.

Here, in the minimum distance corresponding point list, for example, the kth minimum distance corresponding point is represented as Q _k and S _k , respectively, and these coordinates are represented as Q _k (x _k ′, y _k ′) and S _k (x _k ). , Y _k ) and include the subscript. Note that there may be two or more local feature amounts associated with one local feature point, but in the following description, one local feature amount is associated with one local feature point in order to simplify the description. Only. The local feature values of Q _k and S _k are V _q (k) and V _s (k), respectively. In step S502, m sets of corresponding points are registered in the minimum distance corresponding point list.

  Next, in step S521, it is determined whether m is 3 or more. As a result of this determination, if m is less than 3, the similarity cannot be calculated, so the process proceeds to step S522, and the process ends.

  On the other hand, if m is 3 or more as a result of the determination in step S521, a variable VoteMax representing the final number of votes is initialized to 0 in step S503. In step S504, a variable Count indicating the number of iterations of the similarity calculation process is initialized to zero.

  Next, in step S505, it is determined whether or not the iteration count number Count has exceeded a predetermined maximum number of iterations Rn. As a result of this determination, if the iteration count number Count does not exceed the maximum number of iterations Rn, the process proceeds to step S506, and a variable Vote representing the number of votes is initialized to zero. On the other hand, if it is determined in step S505 that the iteration count Count exceeds the maximum number of iterations Rn, the process proceeds to step S519, the final vote number VoteMax is set as the similarity, and the rotation angle is set. In step S520, the process ends.

Next, in step S507, two sets of coordinates of corresponding point pairs are randomly extracted from the minimum distance corresponding point list. Here, the extracted two sets of coordinates are Q ₁ (x ₁ ′, y ₁ ′), S ₁ (x ₁ , y ₁ ) and Q ₂ (x ₂ ′, y ₂ ′), S ₂ (x ₂ , y ₂ ). In step S508, conversion matrices M and T are calculated. Here, the matrix M is a matrix composed of variables a to d in the following equation (8), and the matrix T is a matrix composed of variables ef.

  Here, in the following description, only the similarity transformation is considered in order to simplify the description. In this case, the above equation (8) is rewritten as the following equation (9).

Further, the variables a, b, e, and f in the equation (9) are represented by coordinates Q ₁ (x ₁ ′, y ₁ ′), S ₁ (x ₁ , y ₁ ), and Q ₂ (x ₂ ′, y ₂ ′). ), S ₂ (x ₂ , y ₂ ) can be used to express the following equations (10) to (13).

  Next, in step S509, the corresponding point selection variable k is initialized to 3 in order to select a point other than the two sets of points randomly extracted from the shortest distance corresponding point list in step S507. Then, in step S510, it is determined whether or not the corresponding point selection variable k exceeds the number m of corresponding points registered in the shortest distance corresponding point list. As a result of the determination, if the number of sets m is exceeded, the process proceeds to step S516. The process of step S516 will be described later.

On the other hand, if the result of determination in step S510 is that the number m of pairs has not been exceeded, one set of new corresponding points is extracted from the minimum distance corresponding point list in step S511. Here, the extracted coordinates are _defined as S _k (x _k , y _k ) and Q _k (x _k ′, y _k ′). When extracting, the additional information is used in the same manner as in step S507.

Next, in step S512, S _k (x _k , y _k ) is converted by equation (9) to obtain coordinates S _k ′ (x _k ″, y _k ”). Thereafter, in step S513, the geometric distance between the coordinates S _k ′ (x _k ″, y _k ″) and the coordinates Q _k (x _k ′, y _k ′) is calculated as the Euclidean distance, and the Euclidean distance is a threshold value. It is determined whether or not Td or less. If the result of this determination is that the Euclidean distance is less than or equal to the threshold Td, the process proceeds to step S514, and after the vote number Vote is incremented, the process proceeds to step S515. On the other hand, if the result of the determination in step S513 is that the Euclidean distance is greater than the threshold Td, the process proceeds to step S515 without doing anything.

  Next, in step S515, the corresponding point selection variable k is incremented, and the process returns to step S510, and the above processing is performed until the corresponding point selection variable k exceeds the number m of corresponding points registered in the shortest distance corresponding point list. repeat.

  Next, the case where the corresponding point selection variable k exceeds the number m of corresponding points registered in the corresponding point list in step S510 will be described. In step S516, the value of the vote number Vote and the value of the final vote number VoteMax are compared, and it is determined whether or not the value of the vote number Vote is larger than the value of the final vote number VoteMax. As a result of this determination, if the value of the vote number Vote is larger than the value of the final vote number VoteMax, the process proceeds to step S517.

In step S517, it is replaced with the value of the final vote number VoteMax in value of the number of votes Vote, replacing the transformation matrix M at that time in the variable M _max. Thus, the rotation angle is determined from the variable M _{max of the} transformation matrix M obtained in the similarity calculation process. In step S518, the iteration count Count is incremented, and the process returns to step S505. On the other hand, as a result of the determination in step S516, if the value of the vote number Vote is equal to or less than the value of the final vote number VoteMax, the process proceeds to step S518, the repeat count number Count is incremented, and the process returns to step S505.

  In the description of FIG. 5, only the similarity transformation is considered. However, other geometric transformations such as affine transformation can be dealt with by obtaining a transformation matrix corresponding to each in step S508. For example, in the case of affine transformation, first, in step S507, the number of coordinates of a pair of corresponding points selected at random is set to 3. Next, in Step S508, using Equation (8) instead of Equation (9), the variables a to f may be obtained using the three corresponding points selected in Step S507 (6 points in total).

  In the description of FIG. 5, the example in which the final vote number VoteMax is output as the similarity in step S519 has been described, but other similarities may be calculated. For example, there is a method in which the number m of pairs of corresponding points registered in the minimum distance corresponding point list created in step S502 is output as the similarity without performing the processing after step S504.

  Further, in this embodiment, the method described in Non-Patent Document 4 is used as the local feature amount, and the method described in Non-Patent Document 5 is a method for collating images based on the comparison of local feature points and local feature amounts. A method using RANSAC has been described. However, other methods may be used as long as the degree of rotation between the two images can be calculated and the method of calculating the rotation angle has a low processing load in the process of calculating the degree of similarity.

  For example, when adopting what is described in Non-Patent Document 2 as a local feature amount, a value for normalizing the direction of the local region for each local feature amount (herein referred to as “main direction”) is used. can get. In this case, in step S403, as in step S502, the difference in the main direction of the corresponding points is calculated when the minimum distance corresponding point list is created, and then a simple statistical analysis is performed on the difference in the main direction. The rotation angle can be estimated without increasing the angle.

  Note that statistical analysis methods include, for example, histogram peak detection and average value calculation. As a similarity calculation method in this case, for example, there is a method in which the number m of pairs of corresponding points registered in the minimum distance corresponding point list is used as the similarity. Of course, the above-described method using RANSAC may be used.

  As described above, in this embodiment, when an image is displayed as a search result, the sample image is displayed after being rotationally normalized with reference to the query image. As a result, it is possible to reduce the load on the user due to mental rotation when checking the search result. In addition, since the rotation angle is calculated in the process of calculating the local feature amount necessary for the search, the rotation normalization based on the query image can be performed without significantly increasing the calculation amount.

  Furthermore, since the necessity of rotation normalization is determined based on whether or not the estimated rotation angle exceeds a predetermined value, the determination criterion is dynamically adjusted according to the resource of the device, and the processing speed of the device It is easy to balance the load reduction due to mental rotation of the user.

  In this embodiment, in step S408, the comparison target image sequence sorted in the order of similarity and subjected to the rotation normalization processing as necessary is output as a comparison result. On the other hand, only the region corresponding to the query image in the comparison target image sequence may be output. Or you may comprise so that the comparison destination image sequence before rotation normalization may be output collectively. At this time, for example, the rotation normalized image is displayed in the sub window.

(Second Embodiment)
In the first embodiment, the rotation normalization process is performed in step S407. However, the rotation normalization process may be performed at another timing if the reduction of the perceptual time accompanying the mental rotation of the user can be realized. May be. In this embodiment, an example in which rotation normalization is performed at different timings will be described. Hereinafter, a configuration example in such a case will be described with reference to FIGS. 6 and 7. Note that the configuration of the image search apparatus according to the present embodiment is the same as that shown in FIG. In the present embodiment, only differences from the first embodiment will be described.

FIG. 6 is a flowchart illustrating an example of a processing procedure for outputting a search result by the local feature point / local feature amount comparison unit 107 in the present embodiment. In the present embodiment, the processing from steps S401 to S405 in FIG. 4 is the same as that in the first embodiment, and a description thereof will be omitted. In the present embodiment, the necessity of rotation normalization and the rotation normalization of the sample image are not performed.
In step S601, the comparison result 108 that is not rotationally normalized is output. In the present embodiment, the necessity of rotation normalization is determined and the rotation normalization of the sample image is performed according to a user instruction.

  Next, a case where the user performs a click operation on the GUI displaying the comparison result 108 will be described as an example. Hereinafter, an image clicked by the user is referred to as a click image.

FIG. 7 is a flowchart illustrating an example of a rotation normalization processing procedure when the display result is clicked by the local feature point / local feature amount comparison unit 107.
First, when an input by a click operation is received from the operation unit 109, the process is started. In step S701, it is determined whether or not rotation normalization of the sample image with respect to the query image is necessary. As a result of the determination, if rotation normalization is necessary, the process proceeds to step S702, and if rotation normalization is not necessary, the process proceeds to step S703.

  Here, the determination criterion similar to that of the first embodiment is used as the rotation normalization criterion. That is, it is determined that rotation normalization is necessary when the estimated rotation angle is ± 60 ° or more. However, a determination criterion other than the criterion of the present embodiment may be adopted as long as it is a reasonable criterion.

  In step S702, the rotation normalization process based on the query image is performed on the sample image using the rotation angle estimated in step S403. In step S703, a sample image (click image) subjected to rotation normalization processing as necessary is output as a comparison result.

  As described above, according to the present embodiment, the sample image that is rotationally normalized based on the query image with the user's click operation as a trigger is displayed. Thereby, when confirming a search result, it becomes possible to reduce a user's load by mental rotation by a user's instruction. In addition, since rotation or independent activity is calculated in the process of calculating the local feature amount necessary for the search, rotation normalization based on the query image can be performed without significantly increasing the calculation amount.

  Furthermore, since it is configured so that the necessity of rotation normalization can be determined based on the estimated rotation angle, the determination criterion is adjusted according to the resource of the device, and the load due to the processing speed of the device and the mental rotation of the user is reduced. It becomes easy to balance with the degree.

  As a click method, for example, the rotation normalization process may be executed when the right button of the mouse is clicked, and the enlargement / full size display that is normally used may be performed when the left button of the mouse is clicked.

  Further, in this embodiment, the click operation is described as an example of the user instruction operation. However, if the rotation normalization necessity determination and the rotation normalization of the sample image are performed according to the user operation instruction, other operations may be performed. The instruction operation may be adopted. For example, a method of inputting an image name as the instruction operation may be used. Or you may comprise so that a rotation normalization process may be performed with respect to the image which the user applied the eyes | visual_axis for the fixed time using the existing eyes | visual_axis detection.

(Third embodiment)
In the first and second embodiments described above, it is configured to determine whether or not the rotation normalization of the sample image to be displayed is necessary based on the rotation angle of the sample image based on the query image. In the present embodiment, an example will be described in which the difference between the estimated resolutions of the sample images based on the query image is used as the determination information as the determination criterion when determining whether rotation normalization is necessary. That is, in this embodiment, the difference in resolution between comparison target images is estimated by paying attention to the scale number 304 in FIG. The specific estimation method will be described below. Note that the configuration of the image search apparatus according to the present embodiment is the same as that shown in FIG. In the present embodiment, only differences from the first embodiment will be described.

In the above-described procedure, first, in step S502, distances between local feature amounts are calculated for all combinations of the local feature amount V _q and the local feature amount V _s . Next, a combination (corresponding point) of the local feature quantity V _q and the local feature quantity V _s so that the calculated distance between the local feature quantities is equal to or smaller than the threshold Tv and becomes the minimum distance is extracted, and the correspondence between them is extracted. A list of points corresponding to the minimum distance is created by registering the points.

Here, when there is a combination (corresponding point) of the local feature quantity V _q and the local feature quantity V _s such that the distance between the local feature quantities is equal to or less than the threshold value Tv and the minimum distance, the respective scales are obtained. The numbers are S _q and S _s . Further, it is assumed that when the image to which the local feature value V _q belongs is used as a query image, the similarity of the image to which the local feature value V _s belongs is calculated.

When the difference in resolution is ΔS = S _s −S _q , if the local feature amount V _q and the local feature amount V _s are correct corresponding points (positive corresponding points), the resolution difference ΔS becomes a constant value. . On the other hand, the resolution difference ΔS at the point where the correspondence is incorrect (miscorresponding point) often does not coincide with the resolution difference ΔS at the positive corresponding point. That is, the resolution difference ΔS is calculated for all corresponding points extracted by the distance between the local feature amounts, and the mode value ΔS _mod of the resolution difference ΔS is obtained, so that the corresponding point having the mode value ΔS _mod is obtained. It can be estimated as a positive corresponding point.

The mode value ΔS _mod represents the resolution difference between the comparative images. When ΔS _mod > 0, the resolution of the image to which the local feature point V _s belongs is larger. When ΔS _mod = 0, there is no difference in resolution between the comparative images, and when ΔS _mod <0, the resolution of the image to which the local feature point V _s belongs is smaller.

In this embodiment, the difference in resolution is used for determining whether rotation normalization is necessary. Here, various criteria can be considered as the rotation normalization determination criteria. In the present embodiment, it is determined that the rotation normalization is unnecessary when the mode value ΔS _mod ≧ 4 of the resolution difference. That is, when the resolution ratio of the sample image to the query image is 4 times or more, rotation normalization is not performed. However, a criterion other than the criterion of the present embodiment may be adopted as long as the criterion is a reasonable criterion. Moreover, you may combine with the rotation normalization criteria using the rotation angle etc. which were demonstrated by 1st and 2nd embodiment. As a combination method of determination, logical product or logical sum may be used.

Further, it may be determined whether or not to use the difference in resolution as a criterion according to the rotation angle. For example, the difference in resolution is added to the criterion only when the rotation angle is within ± 30 °, and the difference in resolution is not added to the criterion in other cases. In this case, when the rotation angle is within ± 30 ° and the mode value ΔS _mod ≧ 4 of the resolution difference, rotation normalization is not performed. On the other hand, when the rotation angle is larger than ± 30 °, rotation normalization is always performed regardless of the difference in resolution of the sample image with respect to the query image.

  As described above, according to the present embodiment, when determining whether rotation normalization is necessary or not, it is configured to be used as a criterion for determining whether or not the difference in resolution between sample images based on a query image is smaller than a predetermined value. This makes it easy to adjust the determination criteria according to the resource of the device and balance the processing speed of the device and the degree of load reduction due to the mental rotation of the user.

(Fourth embodiment)
In the first embodiment, when the search result image is displayed, the image is displayed as a rotation-normalized image based on the query image. On the other hand, when it is possible to reduce the perceived time associated with the mental rotation of the user, a method other than the method of rotationally normalizing the search result image may be used.

  In the present embodiment, the auxiliary image clearly indicates how the corresponding region between the query image and the search result image exists in the image, and in particular, the user is mentally rotated by clearly indicating the rotation direction. A method for reducing the perception time will be described. Note that the configuration of the image search apparatus according to the present embodiment is the same as that shown in FIG. In the present embodiment, only differences from the first embodiment will be described. In the present embodiment, a case where a rectangle is used as an auxiliary image will be described.

  FIG. 8 is a flowchart illustrating an example of a processing procedure for outputting a search result by the local feature point / local feature amount comparison unit 107 in the present embodiment. The difference from FIG. 4 is that the process of step S801 is performed instead of the process of step S403, and the process of step S802 is performed instead of the processes of steps S406 and S407. Other processes are the same as those in FIG.

  In step S <b> 801, the similarity between the query image 101 input to the query image input unit 102 and the sample image in the comparison target image sequence 104 is calculated. At this time, a deformation parameter for a corresponding region between the sample image and the query image is estimated using a value generated in the process of calculating the similarity. Details of the process in step S801 will be described later. In addition, the corresponding region information and the deformation parameter with the query image are associated with the corresponding sample image.

  In step S802, the corresponding region of the query image is drawn as a rectangle on the sample image. The drawing uses the corresponding region information and the deformation parameter with the query image obtained by the processing in step S801. In the present embodiment, the corresponding area information with the query image is a set of corresponding points that satisfy the condition of step S513. The rectangle is a circumscribed rectangle that includes all corresponding points on the query image, and is mapped onto the sample image using a deformation parameter.

  Here, as the rectangular design, the color of the rectangle base generated on the query image is set to a color different from the colors of the other sides, and then the rectangle is mapped. This makes it easy to recognize the vertical and horizontal directions of the rectangle.

Next, details of the processing in step S801 will be described.
FIG. 9 is a flowchart showing an example of a processing procedure for calculating the similarity and the deformation parameter by the local feature point / local feature amount comparison unit 107, which is the detailed processing in step S801. The difference from FIG. 5 is that the process of step S901 is performed instead of step S517, and the process of step S902 is performed instead of step S519. Other processes are the same as those in FIG.

In step S901, the value of the final vote number VoteMax is replaced with the value of the vote number Vote, and the variables M _max and T _max are replaced with the conversion matrices M and T at that time. In step S902, the corresponding region information with the query image, the final vote number VoteMax, and the transformation matrices M _max and T _max are set as the deformation parameters at that time.

  As described above, in the present embodiment, when an image is displayed as a search result, the corresponding region of the query image is drawn as a rectangle on the sample image. As a result, the user can easily recognize the up / down / left / right directions of the rectangle, and the user's load due to mental rotation can be reduced when checking the search result.

  In the present embodiment, the color of the bottom of the rectangle is set to a color different from the colors of the other sides. However, the above method can be used as long as it is a method that facilitates recognition of the rectangle in the vertical and horizontal directions by the user. However, other methods may be used. For example, the colors of all sides of the rectangle may be set to different colors. Alternatively, instead of changing the color, a different line width or line type may be used.

  Further, instead of changing the design of the sides of the rectangle, the deformation parameters may be displayed as text after drawing the rectangle. Furthermore, the text color, line width, and font may be changed according to the degree of deformation. As a result, the user can easily recognize the degree of deformation of the sample image with respect to the query image, and can assist mental rotation when checking the search result. Therefore, it becomes possible to reduce a user's load.

  In this embodiment, when displaying the search result, the region corresponding to the query image is drawn in a rectangle on the sample image. However, if the region corresponding to the query image on the sample image can be easily confirmed, You may comprise so that it may draw using another auxiliary image. For example, instead of a rectangle, a circle or an ellipse including all corresponding points on the query image may be used, or an arbitrary polygon may be used. Furthermore, an arrow may be used. However, in that case, in order to apply the above-described method for facilitating the recognition in the vertical and horizontal directions, it is necessary to set a range representing the vertical and horizontal directions in the auxiliary image. Furthermore, on the query image, the corresponding region may be indicated by an auxiliary image having the same or similar design as the auxiliary image on the sample image.

(Other embodiments)
The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, or the like) of the system or apparatus reads the program. It is a process to be executed.

103 First local feature point / local feature quantity extraction unit 106 Second local feature point / local feature quantity extraction unit 107 Local feature point / local feature quantity comparison unit

Claims (12)

Local feature extraction means for extracting local local features from the input image and the comparison image;
Collation means for collating local features of the input image and comparison image extracted by the local feature extraction means to calculate a similarity;
A rotation angle calculation unit that calculates a rotation angle of the comparison image with respect to the input image based on a value obtained when the similarity is calculated by the matching unit;
Rotating means for rotating the comparison image by the rotation angle calculated by the rotation angle calculating means;
An image search apparatus comprising: output means for outputting a comparison image rotated by the rotating means in accordance with the similarity calculated by the collating means. The output means outputs a comparison image not rotated by the rotation means;
The image search apparatus according to claim 1, wherein the rotation unit rotates the comparison image that has received the instruction in response to an instruction for the outputted non-rotated comparison image.   The image search apparatus according to claim 1, wherein the rotation unit rotates a comparison image in which the rotation angle is larger than a predetermined value. The local feature extraction means further extracts the resolution of the input image and the comparison image,
The image search apparatus according to claim 1, wherein the rotation unit rotates the comparison image when a difference in resolution between the input image and the comparison image is smaller than a predetermined value. The local feature extraction means further extracts the resolution of the input image and the comparison image,
The rotating means compares the rotation angle when the rotation angle is larger than a predetermined value, or when the rotation angle is equal to or smaller than the predetermined value and the difference in resolution between the input image and the comparison image is smaller than a predetermined value. The image search apparatus according to claim 1, wherein the image is rotated.   The image search apparatus according to claim 1, wherein the output unit outputs an area corresponding to the input image in the comparison image. Local feature extraction means for extracting local local features from the input image and the comparison image;
Collation means for collating local features of the input image and comparison image extracted by the local feature extraction means to calculate a similarity;
Output means for outputting the comparison image according to the similarity calculated by the matching means,
The output means outputs a comparison image to which an auxiliary image indicating a region corresponding to the input image is added, and the auxiliary image has a design that differs depending on a degree of deformation of the corresponding region. apparatus.   The image search apparatus according to claim 7, wherein the output unit outputs an input image to which an auxiliary image indicating an area corresponding to the comparison image is added. A local feature extraction step of extracting local local features from the input image and the comparison image;
A collation step of calculating a similarity by collating local features of the input image and the comparison image extracted in the local feature extraction step;
A rotation angle calculation step of calculating a rotation angle of the comparison image with respect to the input image based on a value obtained when the similarity is calculated in the collation step;
A rotation step of rotating the comparison image by the rotation angle calculated in the rotation angle calculation step;
And an output step of outputting the comparison image rotated in the rotation step according to the similarity calculated in the collation step. A local feature extraction step of extracting local local features from the input image and the comparison image;
A collation step of calculating a similarity by collating local features of the input image and the comparison image extracted in the local feature extraction step;
An output step of outputting the comparison image according to the similarity calculated in the collation step,
In the output step, a comparison image to which an auxiliary image indicating a region corresponding to the input image is added is output, and the auxiliary image has a different design according to the degree of deformation of the corresponding region. retrieval method. A local feature extraction step of extracting local local features from the input image and the comparison image;
A collation step of calculating a similarity by collating local features of the input image and the comparison image extracted in the local feature extraction step;
A rotation angle calculation step of calculating a rotation angle of the comparison image with respect to the input image based on a value obtained when the similarity is calculated in the collation step;
A rotation step of rotating the comparison image by the rotation angle calculated in the rotation angle calculation step;
A program for causing a computer to execute an output step of outputting a comparison image rotated in the rotation step according to the similarity calculated in the collation step. A local feature extraction step of extracting local local features from the input image and the comparison image;
A collation step of calculating a similarity by collating local features of the input image and the comparison image extracted in the local feature extraction step;
Causing the computer to execute an output step of outputting the comparison image according to the similarity calculated in the matching step;
In the output step, a comparison image to which an auxiliary image indicating a region corresponding to the input image is output is output, and the auxiliary image has a design that varies depending on the degree of deformation of the corresponding region. .

Images