JP5304553B2 - Image processing apparatus and image processing program - Google Patents

Description

  The present invention relates to an image processing apparatus and an image processing program.

  Conventionally, in order to determine the similarity of two documents, for example, conventionally, a document is converted into a raster image, a feature amount of each image is obtained, a similarity conversion parameter is estimated with reference to each feature amount, and one image is obtained. There has been proposed a technique for collating with the other image after scaling with the estimated similarity transformation parameter.

Japanese Patent No. 4140221 Japanese Patent No. 3693147 Japanese Patent Laid-Open No. 10-269357

  An object of the present invention is to improve the estimation accuracy of similarity transformation parameters obtained in the process of judging the similarity of two images.

  The image processing apparatus according to the present invention accumulates pixel values of respective images to be compared in at least one of horizontal and vertical directions, and acquires feature amount information including information for specifying a peak position of the accumulated value. Similarity for superimposing the peak position of one image to the peak position of the other image by causing the peak position of one image to be scaled or translated from the reference point from the feature amount information corresponding to each acquired image A similarity conversion parameter estimating means for estimating a similarity conversion parameter represented by a scaling factor and a movement amount, and an index indicating a similarity relationship between the images after correcting the one image with the similarity conversion parameter; And the similarity transformation parameter estimation means calculates each of the peak positions of one image and the peak position of the other image. Generate a linear function on the scale-moving plane coordinate expressed by the scaling ratio and the moving amount for superimposing each, and the scaling factor-moving-plane coordinate on which similar transformation parameters will exist The predetermined range is divided into a plurality of regions, the number of linear functions passing through each region is obtained, and the set of peak positions of each image that has generated each linear function is specified for each region. If there are areas where the linear functions based on the same peak position overlap in the set of specified peak positions, the area is subdivided into a plurality of areas until there is no overlapping area. Then, the number of linear functions that pass through each of the subdivided areas is calculated and the process of specifying the set of peak positions of each image that has generated each linear function is repeated, and the number of linear functions that pass through is maximized. The area Request similarity transformation parameter based on a linear function over to, characterized in that.

  Further, the similarity transformation parameter estimation means detects a region where the number of passes through the linear function is maximum, and then targets only the detected region to the same peak position in the specified set of peak positions. It is characterized in that a region where a linear function based thereon passes through is detected.

  Further, when the similarity transformation parameter estimation means subdivides the region into a plurality of regions, the intersection of the linear function is obtained in the region to be divided, and the obtained intersection exists among the regions generated by the subdivision. Only the region to be performed and the region adjacent to the region are targeted for obtaining the number of linear functions passing therethrough and specifying the set of peak positions of each image in which each linear function is generated.

  The image processing program according to the present invention includes feature amount information including information for accumulating pixel values of images to be compared in at least one of horizontal and vertical directions and specifying a peak position of the accumulated value. And a feature amount information corresponding to each acquired image to superimpose the peak position of one image on the peak position of the other image by performing at least one of enlargement / reduction or translation from the reference point A similarity transformation parameter estimating means for estimating a similarity transformation parameter represented by an enlargement / reduction ratio and a movement amount, and correcting the one image with the similarity transformation parameter, and then showing a similarity relationship between the images. Functioning as a means for calculating an index, wherein the similarity transformation parameter estimation means determines each peak position of one image as the other Generate a linear function on the plane coordinate of the scaling factor-moving amount expressed by the scaling factor and the moving amount to be superimposed on each of the peak positions of the image, and the scaling factor-moving amount that the similarity conversion parameter will exist A predetermined range on the plane coordinates is divided into a plurality of regions, the number of linear functions passing through each region is obtained, and the set of peak positions of each image that has generated each linear function is specified for each region. If there is a region where the linear function based on the same peak position passes through in the set of peak positions specified for each region, a plurality of such regions are excluded until there is no region where the overlapping passes. Subdividing into regions, calculating the number of linear functions that pass through each of the subdivided regions, and repeating the process of specifying the set of peak positions of each image that generated each linear function, and passing the linear function That number seek similarity transformation parameter based on a linear function passing through a region to be a maximum, it is characterized.

  According to the first and fourth aspects of the invention, it is possible to improve the estimation accuracy of the similarity transformation parameter obtained in the process of judging the similarity between two images, compared to the case where the present configuration is not provided. .

  According to the second and third aspects of the present invention, the similarity determination between two images is realized in a shorter time or with a smaller memory usage compared to the case where the target area is not specified. be able to.

1 is a block diagram illustrating an image processing apparatus according to an embodiment of the present invention. FIG. 2 is a hardware configuration diagram of a computer forming an image processing apparatus according to the present embodiment. It is the flowchart which showed the image collation process in this Embodiment. It is a figure used in order to demonstrate the production | generation of the projection waveform and the feature-value of a document image in this Embodiment. It is the flowchart which showed the fluctuation | variation correction parameter estimation process in this Embodiment. It is the figure shown about matching with the local peak coordinate in a reference document image, and the local peak coordinate in an input document image in this Embodiment. It is the figure which showed an example of the sk plane coordinate produced | generated in this Embodiment. It is the figure which showed an example of the sk plane coordinate when the inside of the correction range set to the sk plane coordinate shown in FIG. 7 is divided | segmented into several cells. It is the figure shown about the positional relationship of the local peak coordinate in the reference document image in which a duplicate vote may generate | occur | produce in this Embodiment, and the local peak coordinate in an input document image. It is the figure which showed an example of the cell in which the duplicate vote has generate | occur | produced in this Embodiment. It is the figure which showed the state when the cell shown in FIG. 10 is subdivided. It is the figure which showed the part where a linear type crosses on sk plane coordinate in this Embodiment. It is the figure which showed the principal part of the said cell when a cell is subdivided in this Embodiment.

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

  FIG. 1 is a block configuration diagram showing an embodiment of an image processing apparatus 20 according to the present invention, and FIG. 2 is a hardware configuration diagram of a computer forming the image processing apparatus 20. The computer forming the image processing apparatus 20 in the present embodiment can be realized with a general-purpose hardware configuration that has existed in the past. That is, as shown in FIG. 2, the computer is provided with a CPU 1, a ROM 2, a RAM 3, an HDD controller 5 connected to a hard disk drive (HDD) 4, a mouse 6 and a keyboard 7 provided as input means, and a display device. An input / output controller 9 for connecting each display 8 and a network controller 10 provided as a communication means are connected to an internal bus 11.

  Returning to FIG. 1, the image processing apparatus 20 includes projection waveform generation units 21 and 22, feature amount generation units 23 and 24, a fluctuation correction parameter estimation unit 25, a projection waveform correction unit 26, and a matching score calculation unit 27. . The projection waveform generation unit 21 generates a projection waveform from the input document image. The projection waveform generation unit 22 generates a projection waveform from the reference document image. The projection waveform generation unit 21 and the projection waveform generation unit 22 may be the same in the process to be performed except that the images to be processed are different. The feature amount generation unit 23 generates a feature amount of the input document image from the projection waveform generated by the projection waveform generation unit 21. The feature amount generation unit 24 generates a feature amount of the reference document image from the projection waveform generated by the projection waveform generation unit 22. The feature quantity generation unit 23 and the feature quantity generation unit 24 may be the same in the process to be performed except that the projection waveform of the image to be processed is different. When the variation correction parameter estimation unit 25 acquires the feature amount of each document image obtained by each of the feature amount generation units 23 and 24, the variation correction parameter estimation unit 25 estimates the variation correction parameter as a similarity conversion parameter from the feature amount of each document image. The projection waveform correction unit 26 corrects the projection waveform of the input document image with the variation correction parameter obtained by the variation correction parameter estimation unit 25. The collation score calculation unit 27 collates the reference document image with the projection waveform of the input document image corrected by the projection waveform correction unit 26 to thereby obtain a collation score serving as a determination index for the similarity relationship between the reference document and the input document. calculate.

  Each component 21 to 27 in the image processing device 20 is realized by a cooperative operation of a computer that forms the image processing device 20 and a program that operates on the CPU 1 mounted on the computer.

  Further, the program used in this embodiment can be provided not only by communication means but also by storing it in a computer-readable recording medium such as a CD-ROM or DVD-ROM. The program provided from the communication means or the recording medium is installed in the computer, and various processes are realized by the CPU of the computer sequentially executing the installation program.

  A characteristic feature of the present embodiment is that a duplicate vote is detected in the fluctuation correction parameter estimation process in the fluctuation correction parameter estimation unit 25, and a duplicate vote is detected when a duplicate vote is detected. The cell is subdivided to eliminate the duplicate vote, and the fluctuation correction parameter is estimated based on the cell having the maximum vote value when the duplicate vote is eliminated. Other processing may be the same as the processing content of Patent Document 1.

  Next, image collation processing in the present embodiment will be described with reference to the flowchart shown in FIG.

  When the images of the input document and reference document to be compared are sent, the projection waveform generation units 21 and 22 receive the respective document images (step 100), and the projection waveforms in the X and Y directions from each image. Is generated (step 200). Then, the feature amount generation units 23 and 24 generate feature amounts of the respective images based on the generated projection waveforms (step 300). The processing so far will be described in detail with reference to FIG.

  Note that the image of the reference document is rasterized image data of the document that is referred to when collating with the image of the input document. The image of the input document is rasterized image data of the document that is collated with the image of the reference document. That is, the image processing apparatus 20 according to the present embodiment provides a collation score as an index for determining whether or not the input document is similar to the reference document. The image data of each document designated by the user is read from the image database and input to the image processing apparatus 20 or read by an image input means such as a scanner and converted into raster image data. Or input after In the projection waveform and feature amount generation processing described here, the same processing is performed on each image of the input document and the reference document. Therefore, here, the image of the input document will be described as a representative.

  When an image of the input document is input, the projection waveform generation unit 21 forms a projection waveform for at least one of the horizontal direction and the vertical direction with respect to the raster image data. For example, if it is known in advance whether the input document is vertical writing or horizontal writing, and it is known that the characteristics of individual documents can be sufficiently expressed only by the projection waveform in the direction, only one direction can be obtained. It may be only necessary to form a projection on the screen. However, if the characteristics of the document image data are unclear, it is preferable to form a projection waveform in both directions. In the present embodiment, a projection waveform is generated in both the X and Y directions.

  Here, the projection refers to a process of accumulating pixel values of pixels located on a certain path along the X direction or the Y direction with respect to raster-format image data. When there are a plurality of paths along the X direction or the Y direction, a sequence of the accumulated results forms a projection waveform. The projection waveform generation unit 21 obtains an accumulation result of pixel values for each path. In this case, practically, it is preferable to correct a slight inclination at the time of reading using a known inclination correction technique before the projection waveform forming process.

  In the case of a path in the X direction, the projection waveform generation unit 21 generates a projection waveform by accumulating pixel values of pixels on the path in the X direction for the document image 12. In this embodiment, a position representing a characteristic portion that satisfies a predetermined condition is selected from the route. As an example of a predetermined condition for selecting a position representing a characteristic portion, there are a differential waveform feature amount such as an extreme value (local peak) and an inflection point. In the present embodiment, the position that becomes the local peak value of the accumulated value is selected. Then, the feature value generation unit 23 generates the feature value (xn, pxn) of the position by combining the coordinate (xn) of the position that becomes the selected local peak and the projection peak value (pxn) of the position. Then, a list of these feature amounts is generated as an input document image feature amount in the X direction. Similarly, in the Y direction, the coordinates (ym) of the position that becomes the local peak and the projection peak value (pym) of the position are generated as a pair to generate the feature amount (ym, pym) of the position. Is generated as an input document image feature amount in the Y direction. In this embodiment, since the local peak coordinates are extracted in the X and Y directions, the values of n and m do not always match.

  The projection waveform generation unit 22 and the feature amount generation unit 24 perform the same processing to generate a projection waveform and reference document image feature amounts in the X and Y directions. Note that the numbers of local peak coordinates in the X direction of the input document image and the reference document image do not always match. The same applies to the Y direction.

  For the reference document, the process up to the feature value generation process may be performed in advance, the process result may be stored in the HDD 4, and the process result value may be read and used in subsequent processes.

  Subsequently, the fluctuation correction parameter estimation unit 25 uses the fluctuation correction parameters such that the projected waveforms in the X and Y directions of the two documents most overlap from the document image feature amounts of the reference document image and the input document image generated by the above processing. Is estimated (step 400). When a plurality of sets of fluctuation correction parameters are estimated, a list of fluctuation correction parameters is generated. The fluctuation correction parameter estimation unit 25 performs processing so that the local peak coordinates of the input document image are superimposed on the local peak coordinates of the reference document image in order to collate the reference document with the input document. The document image is enlarged or reduced, or moved in parallel. That is, the fluctuation correction parameter is represented by the scaling factor and the movement amount of the input document image when the local peak coordinates of the input document image are superimposed on the local peak coordinates of the reference document image in each of the X and Y directions. . The fluctuation correction parameter estimation processing in the fluctuation correction parameter estimation unit 25 will be described with reference to the flowchart shown in FIG.

  The fluctuation correction parameter estimation unit 25 performs the following processing for each feature amount of each image of the reference document and the input document for each of the X and Y directions. First, in the X direction, as illustrated in FIG. 6, m local peak coordinates a1 to am are included in the feature amount of the reference document image, and n local peak coordinates b1 to bn are included in the feature amount of the input document image. If included, the fluctuation correction parameter estimation unit 25 calculates the local peak coordinates (ai (1 ≦ i ≦ m)) from the reference document image feature value, and the local peak coordinates (bj (1 ≦ 1) from the input document image feature value. j ≦ n)) one by one, and a straight line indicating the relationship between the scaling ratio and the movement amount when the local peak coordinates bj in the input document image feature quantity are superimposed on the local peak coordinates ai in the reference document image feature quantity Generate an expression. This linear equation can be expressed by a linear function equation of ai = k × bj + s on the s-k plane coordinates. In the equation, k is an enlargement / reduction ratio when the projection waveform is enlarged / reduced in order to superimpose bj on ai with the origin as a reference point, and s is an amount of movement when bj is translated along with enlargement / reduction. A list of linear expressions is generated by generating similar linear expressions for all combinations of local peak coordinates (step 401). That is, the list of linear expressions includes m × n linear expressions. FIG. 7 shows an example in which this linear expression is drawn on the sk plane coordinates. When generating a straight line list, the fluctuation correction parameter estimating unit 25 stores a set of local peak coordinates (ai and bj) associated with generating a straight line expression in association with the straight line expression.

  The above processing is also performed in the Y direction. Through the processing so far, the s-k plane coordinates illustrated in FIG. 7 are generated in the X and Y directions, respectively.

  By the way, regarding the expansion / contraction rate k, k = 1 means no expansion / contraction, 0 <k <1 means reduction, and k> 1 means enlargement. For example, a range in which the expansion / contraction rate k, which is a fluctuation correction parameter, will exist, The maximum value kmax = 10.0 and the minimum value kmin = 0.1 are determined in advance. Similarly, for the movement amount s, for example, the movement range of the movement amount s is determined in advance as a maximum value smax = 50 pixels, a minimum value smin = -50 pixels of the movement range, and the like. Then, the preset correction range is divided into a predetermined number such as 16 divisions (step 402). The small area obtained by this division is called a cell. An example of the sk plane coordinates when the correction range of the sk plane coordinates is divided into a plurality of parts is shown in FIG. The predetermined number for dividing the correction range will be described later.

  Subsequently, the fluctuation correction parameter estimation unit 25 votes a straight line for each cell (step 403), and detects a cell having the maximum vote value (step 404). Voting means adding up the number of straight lines that pass through the cell. When a plurality of cells having the maximum vote value are detected, a list of those cells is generated.

  Then, the fluctuation correction parameter estimation unit 25 detects a cell that becomes a duplicate vote from the cells that have the largest vote (step 405). Here, the duplicate vote will be described.

  For example, as illustrated in FIG. 9, the local peak coordinates b2 and b3 included in the input document image feature are slightly enlarged or reduced, or moved slightly, so that the local peak included in the reference document image feature is obtained. It is assumed that it is overlapped with the coordinates a2. FIG. 10 is a diagram showing a certain cell through which a linear expression passes, but a linear expression generated based on the local peak coordinates a2 and b2 and a straight line generated based on the local peak coordinates a2 and b3 as in this example. Like a formula, the straight line formula generated by the same local peak coordinate a2 may pass through the same cell redundantly. Such a case may occur when ruled lines drawn at intervals of several pixels are included in the document image. However, the local peak coordinates a2 are not overlapped simultaneously with the local peak coordinates b2 and b3. That is, one local peak coordinate cannot be associated with a plurality of local peak coordinates at the same time. In the present embodiment, the fact that a linear expression generated based on one local peak coordinate is voted for one cell is referred to as a duplicate vote, but a cell including such a duplicate vote. The number of votes in is not correct. In step 406, a cell in which duplicate voting illustrated in FIG. 10 is performed is detected.

  The fluctuation correction parameter estimation unit 25 detects a duplicate vote by referring to local peak coordinates associated with each linear expression passing through the cell when generating a list of linear expressions. If no duplicate vote is detected (N in step 406), the cell having the maximum vote value is specified (step 407). There may be a case where a plurality of cells having the maximum vote value are specified. Then, the fluctuation correction parameter estimation unit 25 estimates the fluctuation correction parameter (step 408). This process will be described in detail later.

  On the other hand, when a duplicate vote is detected (Y in step 406), the fluctuation correction parameter estimation unit 25 subdivides only the cells in which the duplicate vote is detected into a predetermined number (step 409). Since only cells in which duplicate voting is detected are targeted for re-division, other cells need not be loaded into the memory (RAM 3) for re-division. FIG. 11 shows an example when the cell illustrated in FIG. 10 is subdivided. When duplicate voting is detected in a plurality of cells, re-division is performed in all the detected plurality of cells. The predetermined number to be subdivided will be described later. Then, a new vote is performed in the re-divided cell (step 410). After voting for each subdivided cell, the process returns to the process of detecting the cell having the maximum vote value (step 404). The process of subdividing and voting cells (steps 409 and 410) is recursively repeated until no duplicate voting is detected.

  As described above, when a cell having the maximum vote value is identified in a state where there is no duplicate voted cell, the fluctuation correction parameter estimation unit 25 performs scaling of the intersection of the linear expressions included in the cell, that is, the fluctuation correction parameter. The rate k and the movement amount s are estimated (step 408).

  When the input document and the reference document are in a similar relationship, only one set of the scaling ratio k and the movement amount s is basically obtained even when a plurality of cells having the maximum vote value are specified. In other words, even if there is no intersection point in the cell that has the maximum vote value, the intersection point that is obtained from the linear expression that passes through the cell matches the intersection point in the cell that has the other maximum vote value. Should do. However, theoretically, a plurality of intersections may be obtained due to factors such as the reading accuracy of the scanner. In this case, a plurality of variation correction parameters are estimated. Here, the process of estimating the fluctuation correction parameter (step 408) will be described in detail.

  If it is assumed that the reference document and the input document are in a similar relationship, theoretically, there is only one set of the correct scaling factor k and the movement amount s. It should intersect at an intersection (s *, k *) represented by the expansion / contraction ratio k of the plane coordinates and the movement amount s. However, in practice, some errors occur due to the reading accuracy of the scanner, the expansion and contraction of the paper of each document, the influence of noise, and the like, and this is where the linear equations essentially intersect at the intersection (s *, k *). As shown in FIG. 12, there may be a slight deviation. Therefore, in the present embodiment, as described below, each parameter value of the scaling factor k and the movement amount s is corrected to obtain one intersection (s *, k *) from a plurality of straight lines.

という式にて表せる。また、セル内を通過する直線群から最小距離にある点の座標を（ｓ＊，ｋ＊）とすると、この点は各直線からの距離の二乗和が最小となる点として求める。つまり、

という式にて表せる。但し、上式におけるＣは、セル内を通過する直線群のパラメータ添字（ｉ，ｊ）を表す。このような（ｓ＊，ｋ＊）は次の式を解くことにより、解析的に求める。

In the present embodiment, the coordinates of the point at the minimum distance from the straight line group are regarded as the coordinates of the intersection. That is, if the distance from the point (s, k) in the cell to the straight line ai = k × bj + s is d _{i, j} ,

It can be expressed by the formula If the coordinates of a point at a minimum distance from a group of straight lines passing through the cell are (s *, k *), this point is obtained as a point where the sum of squares of the distance from each line is minimum. That means

It can be expressed by the formula However, C in the above formula represents a parameter subscript (i, j) of a straight line group passing through the cell. Such (s *, k *) is obtained analytically by solving the following equation.

  As described above, the fluctuation correction parameter estimation unit 25 obtains the point (s *, k *) in the cell, that is, the expansion / contraction rate k and the movement amount s as the fluctuation correction parameters.

  In the above description, it is assumed that voting is performed again for all the cells re-divided in step 410. However, if subdivision is performed and cell A is considered to have an intersection as shown in FIG. 13, even if voting is performed on cell C in the same manner as cell A, the intersection obtained from the linear expression of cell C is Since it exists in A, the voting process is useless. Therefore, when the cell is subdivided, the cell A in which the intersection exists among the subdivided cells is obtained before the voting in step 410, and the cell C that is not adjacent to the cell A is processed after the voting or the like. It may be excluded. That is, the data relating to the cell C need not be loaded into the memory. Note that the cell B adjacent to the cell A may be excluded from the processing target, but the intersection that should exist in the cell B is actually erroneously estimated to exist in the cell A due to an error. In consideration of the case, the adjacent cell B may not be excluded from processing.

  Here, the correction range and cell division will be described.

  As described above, in the present embodiment, the correction range is divided into a plurality of cells in step 402. Here, if the correction range is roughly divided (decreasing the number of cells), the number of cells in which duplicate voting occurs increases, and the load on processing for voting to each area after subdivision and subdivision Will increase. On the other hand, if the correction range is finely divided (increasing the number of cells), the processing load required for detecting a linear vote or a duplicate vote increases. As described above, the number of divisions of the correction range is in a trade-off relationship. However, in the present embodiment, when a duplicate vote occurs, these problems can be solved by subjecting only the generated cell to the subdivision and re-voting. In other words, the predetermined number of correction ranges to be divided has a high degree of freedom, and even if it is set appropriately, improvement in the processing speed and memory usage can be expected compared to the conventional case. The same applies to cell subdivision.

  Returning to FIG. 3, when the fluctuation correction parameter is obtained as described above, the projection waveform correction unit 26 inputs the input document with the fluctuation correction parameter (scale rate k and movement amount s) obtained by the fluctuation correction parameter estimation unit 25. The projected waveform of the image is corrected (step 500). When a plurality of variation correction parameters are obtained, the projection waveform of the input document image is corrected with each variation correction parameter.

  The collation score calculation unit 27 becomes a determination index of the similarity between the reference document and the input document by collating the projection waveform of the reference document image with the projection waveform of the input document image corrected by the projection waveform correction unit 26. A verification score is calculated (step 600). The matching score may be obtained for each of the X and Y directions by using a correlation coefficient, for example.

  By referring to the matching score calculated by the matching score calculation unit 27, the operator determines the similarity between the input document and the reference document. Alternatively, the presence or absence of a similarity relationship is determined by comparing the collation score with a preset threshold value. When a plurality of sets of collation scores are calculated because a plurality of fluctuation correction parameters have been obtained, only the one having a larger collation score may be recognized as a correct fluctuation correction parameter for determination.

  By the way, when the collation score is the sum of correlation values in the X and Y directions, the correlation values in the X and Y directions are 0.7, 0.7, 0.4, and 1.0. The matching score is calculated to be 1.4. However, in the latter case, it may be determined that there is a similarity between the reference document and the input document, regardless of whether the correlation value in the X direction is as low as 0.4. Therefore, if the previously obtained correlation value is smaller than a predetermined threshold value, it may be determined that there is no similarity between the input document and the reference document, and the subsequent processing may be skipped.

  By the way, the similarity relationship between the input document and the reference document may be determined without operating the matching score calculation unit 27. For example, when an input document is enlarged / reduced, the enlargement / reduction ratios in the horizontal direction and the vertical direction should normally be enlarged / reduced at a substantially equal ratio. That is, normally, copying is performed at the same vertical / horizontal magnification, and copying is not performed extremely vertically or horizontally. In the present embodiment, the fluctuation correction parameter estimation unit 25 estimates the fluctuation correction parameters (the enlargement / reduction ratio k and the movement amount s) in the X and Y directions, but the X direction and the Y direction in any document image. When the fluctuation correction parameter value of the input document image is too different, for example, when the enlargement / reduction ratio in the X direction in the input document image is not in the range of 0.9 to 1.1 times the Y direction, the difference is more than a predetermined difference. Therefore, the subsequent processing is skipped because an appropriate fluctuation correction parameter is not obtained. That is, when the scaling ratios k in the X and Y directions are different from each other by a predetermined value or more, it is determined that there is no similarity between the input document and the reference document without causing the matching score calculation unit 27 to calculate the matching score.

  1 CPU, 2 ROM, 3 RAM, 4 hard disk drive (HDD), 5 HDD controller, 6 mouse, 7 keyboard, 8 display, 9 input / output controller, 10 network controller, 11 internal bus, 20 image processing device, 21 and 22 Projection waveform generation unit, 23, 24 Feature amount generation unit, 25 Variation correction parameter estimation unit, 26 Projection waveform correction unit, 27 Collation score calculation unit.

Claims (4)

Means for accumulating pixel values of each image to be compared in at least one of the horizontal and vertical directions and acquiring feature amount information including information for specifying a peak position of the accumulated value;
Similarity conversion parameters for superimposing the peak position of one image on the other image by performing at least one of enlargement / reduction or translation from the reference point, based on the acquired feature amount information corresponding to each image A similarity conversion parameter estimating means for estimating a similarity conversion parameter represented by a scaling ratio and a movement amount;
Means for calculating an index indicating a similarity relationship between the images after correcting the one image with the similarity transformation parameter;
Have
The similarity transformation parameter estimation means includes
Generating a linear function on a plane coordinate of an enlargement / reduction ratio-movement amount represented by an enlargement / reduction ratio and a movement amount for superimposing each of the peak positions of one image with each of the peak positions of the other image;
Dividing a predetermined range on the plane coordinate of the scaling ratio-movement amount plane where the similarity conversion parameter will exist, into a plurality of regions,
Determine the number of linear functions that pass through each area and identify the set of peak positions of each image that generated each linear function for each area,
If there is a region where the linear functions based on the same peak position overlap in the set of peak positions specified for each region, the region is divided into a plurality of regions until there is no overlapping region. Re-divided into two, repeatedly calculating the number of linear functions that pass through each of the subdivided regions and identifying the set of peak positions of each image that generated the respective linear functions,
Find similarity transformation parameters based on a linear function that passes through the region where the number of linear functions passes is maximum,
An image processing apparatus. The similarity transformation parameter estimation means includes
A region where a linear function based on the same peak position is duplicated and passed through only the detected region after detecting a region where the number of passes through the linear function is maximum. The image processing apparatus according to claim 1, wherein the detection is performed. The similarity transformation parameter estimation means includes
When the area is subdivided into multiple areas, the intersection of the linear function is obtained in the area to be divided, and among the areas generated by the subdivision, only the area where the obtained intersection exists and the area adjacent to the area The image processing apparatus according to claim 1, wherein the number of linear functions that pass is determined and the set of peak positions of each image that has generated each linear function is specified. Computer
Means for accumulating pixel values of each image to be compared in at least one of the horizontal and vertical directions and acquiring feature amount information including information for specifying a peak position of the accumulated value;
Similarity conversion parameters for superimposing the peak position of one image on the other image by performing at least one of enlargement / reduction or translation from the reference point, based on the acquired feature amount information corresponding to each image A similarity conversion parameter estimating means for estimating a similarity conversion parameter represented by a scaling ratio and a movement amount;
Means for calculating an index indicating a similarity relationship between the images after correcting the one image with the similarity transformation parameter;
Function as
The similarity transformation parameter estimation means includes
Generating a linear function on a plane coordinate of an enlargement / reduction ratio-movement amount represented by an enlargement / reduction ratio and a movement amount for superimposing each of the peak positions of one image with each of the peak positions of the other image;
Dividing a predetermined range on the plane coordinate of the scaling ratio-movement amount plane where the similarity conversion parameter will exist, into a plurality of regions,
Determine the number of linear functions that pass through each area and identify the set of peak positions of each image that generated each linear function for each area,
If there is a region where the linear functions based on the same peak position overlap in the set of peak positions specified for each region, the region is divided into a plurality of regions until there is no overlapping region. Re-divided into two, repeatedly calculating the number of linear functions that pass through each of the subdivided regions and identifying the set of peak positions of each image that generated the respective linear functions,
Find similarity transformation parameters based on a linear function that passes through the region where the number of linear functions passes is maximum,
An image processing program characterized by that.

Images