JPH07141504A - Image recognition device - Google Patents

Abstract

PURPOSE:To recognize even an object whose size and attitude are unknown on condition that it is the same model and to correctly recognize even an object in a complicated background. CONSTITUTION:A conversion part 12 which converts a series of evaluation points into a similar model or rotated model is added to a conventional recognition device which recognizes an image by contrasting directional features extracted from an input image 2 by parallel arithmetic with a model 1 represented as a series of evaluation points, and then the object 2 whose size and attitude are unknown can be recognized. Further, the device is provided with means (8, 9, etc.) which extracts nondirectional features from the model 1 to evade misrecognition due to strongly directional features extracted from the complicated background, thereby enabling correct recognition.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition apparatus for extracting a desired model feature from an image based on its position and degree of matching by performing local parallel processing on an input image.

[0002]

2. Description of the Related Art Applicants first expressed a model to be recognized as a series of evaluation points, and for each evaluation point, the value of the directional feature surface to be referred to and the amount of movement from the previous evaluation point. Is calculated in the dictionary table, and the operation of moving the counter surface and adding the value of the specified input direction feature surface to the counter surface is performed for all evaluation points, and then the local maximum value of the counter surface ( We have proposed a method for recognizing an arbitrary shape including a symbol by finding the maximum value (Japanese Patent Application No.
-198853). In addition, as a method of distinguishing between similar models, a difference pattern is extracted from similar models, and in addition to the conventional evaluation point series (positive evaluation point), the evaluation point series (negative evaluation point) of the difference pattern part is also taken into consideration. We have proposed a method that enables discrimination of similar models by adding the reference direction characteristic surface to the counter surface at the evaluation point, and subtracting the reference direction characteristic surface from the counter surface at the negative evaluation point. See Japanese Patent Application No. 2-90441).

[0003]

However, when the recognition target causes an error that cannot be absorbed due to an elastic point or the like, or when the recognition target has a similarity variation or a posture variation accompanied by rotation, recognition cannot be performed by the conventional method. There is a problem. In addition, generally known template matching cannot handle such errors and fluctuations.
It is necessary to separately prepare a model for each size and posture, and it is necessary to prepare a large number of models in order to cope with such deformation, which is a problem that it becomes expensive. In addition, since the background of the image to be recognized is complicated or coarse, strong directional features are detected in the part where the model to be recognized does not exist, and it is recognized in other than the desired position or model. There is also a problem. Therefore, an object of the present invention is to enable recognition with the same model even if the size and orientation of the recognition target are unknown, and to prevent erroneous recognition due to strong directional features extracted from a complicated background part. To avoid it.

[0004]

In order to solve such a problem, according to the first invention, a directional feature extracted from an input image by parallel operation is compared with a model of a predefined shape. In an image recognition device for image recognition, an image input means for inputting a recognition target image and a reference model image, an image storage means for storing the input image, and a model of a shape to be extracted are expressed as a series of evaluation points. However, a value related to the directional feature to be referred to for each evaluation point and a movement amount from the immediately preceding evaluation point in the series of evaluation points are defined, and a value and a movement amount related to the defined directional feature. When the models have a high degree of matching with each other, the value relating to the directional characteristics of the evaluation points corresponding to the difference patterns and the amount of movement are distinguished from the similar model. Addition means for adding to the evaluation point sequence of the model as information for, and direction plane creating means for creating a plurality of direction planes expressing the strength of the direction component by extracting the directional features of the input image, A counter surface storage unit that stores a counter surface for each point of the input image, a moving unit that moves the counter surface in parallel according to the amount of movement defined for each evaluation point, and the moving unit that moves the counter surface. Adder / subtractor for adding or subtracting the data of the directional surface obtained by a value relating to the directional characteristic to a counter surface, and repeating means for repetitively executing the parallel movement and the add / subtract in the order of the evaluation point series of the model. A detecting means for detecting the shape and position of the model from the distribution of the sum of the directional surfaces for each point of the counter surface after the parallel movement and the addition and subtraction are repeated. First conversion means for converting the evaluation point series into a similar model, second conversion means for converting the evaluation point series into a rotated model, and first estimation means for estimating the size of the target image Second estimation means for estimating the rotation direction and the angle of the target image, and are provided. Further, in the second invention, the model of the shape to be extracted in the first invention is expressed as a series of evaluation points, and a value relating to non-direction representing the presence or absence of a model for each evaluation point and the previous one in the series. The feature is that the amount of movement from the evaluation point is defined, and extraction means for extracting these amounts is added, and this is recognized in combination with the evaluation point series having the directional characteristic.

[0005]

When the image is recognized by the degree of matching between the directional feature extracted from the input image by the parallel operation and the model expressed as a series of evaluation points, the center of the model is compared with the evaluation point series of the model. Or the evaluation point, or move the evaluation point to a point on the straight line between the two points having a distance that is a coefficient times the distance between the two points, or
Alternatively, after moving all evaluation points by multiplying the size of the vector between all evaluation points by a factor, the recognition by parallel operation can be performed to deal with errors or similar deformation of the recognition target. To get it. Similarly, when the recognition point changes due to the error of the recognition target or the rotation of the posture, when the evaluation point is on the circumference of the circle with the center of the model as the center of the arc in the evaluation point series of the model, Even if the posture of the recognition target changes due to rotation, by performing the processing by moving the evaluation points to the points on the circumference that have moved by a certain angle between the points for all the evaluation points, the recognition by parallel calculation is performed. It can be dealt with. Furthermore, when the shape of the recognition target, such as a rough outline, is known, the size of the recognition target is estimated, and recognition is performed based on the estimated value, thereby significantly reducing the amount of calculation. Particularly, the posture angle is roughly estimated by a histogram of angles extracted from the angular characteristic surface. In addition, when creating a series of model evaluation points, the conventional evaluation point series with directional characteristics does not have directional characteristics based on points such as the center line of the model that can determine the presence or absence of the model (non-directional Features) By expressing the model by adding a series of evaluation points, the part that is not the original recognition target, which is detected by the complicated background, has strong directional features, and its position in the image is similar to the model. It avoids erroneous recognition that occurs in 1) and enables recognition of a desired position and model.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Before describing the embodiments of the present invention, first, a directional characteristic surface and a model, which form the central concept of the present invention, will be described. At each point of the target image information, the direction characteristic element of the N (8 in this example) plane based on the density gradient,
That is, the direction characteristic surface U is created. U = {u _n (i, j)}, n = 0 to 7 1 ≦ i ≦ I, 1 ≦ j ≦ J (1) Here, the directional feature plane u _n (i, j) is the target image. A point (i, j) separated by a mesh and its neighboring points and FIG.
16 shows the strength of the directivity in the n direction determined by the direction n defined in FIG.
~ Edge operator E0 ₀ to E showing the concept (h)
The output of the _0-7 (however, a negative value is cut)
Alternatively, it can be obtained by binarizing it. For example, the image information is a line in the vertical direction (directions 2 and 6 in FIG. 16), the left side from the boundary is all “1” (black), and the right side is all “0”.
In the case of (white), considering the case of performing calculation of an edge operator to point to the vicinity "1" in the left boundary in the case of n = 3, the operator E0 ₃ as shown in FIG. 17, the point of operation , The product sum operation of each value of the operator E0 ₃ and the image information (“1” or “0”), 0 ×
1 + (-1) x1 + (-1) x0 + 1x1 + 0x1 +
From (−1) × 0 + 1 × 1 + 1 × 1 + 0 × 0 = + 2, u
₃ (i, j) = 2 is obtained. Such calculation is shown in FIG.
This is performed for the eight directional surfaces shown in (A) to (H). When the result of the operation is negative, the value is set to 0. Therefore, the output from this edge operator shows the slope of the density at that point, and the direction is from lower to higher.

[0007] However, only the operator E0 ₀ ~E0 ₇ output, and that feature elements is localized, for the direction accuracy is loose integration with perpendicular direction of the information of the inclination, the linear components in the vicinity thereof It is convenient for future processing to replace it with the feature element that exists. This integration process
The MAP method previously proposed by the inventors (a method of performing contraction / expansion calculation on a directional feature element: if necessary, for example, “Journal of the Institute of Electronics, Information and Communication Engineers, Vol. J74-D-II, No. 6, p.
p. 718-726, June 1991 ”) and the like, and it is possible by repeating the local parallel operation. However, the format of the expression and the meaning of the characteristic element itself do not change even after the integration, even if they are the output of the edge operator. Therefore, in the following description, u _n (i, j) is shown in FIG.
It is assumed that the outputs are those of the edge operators E0 _{0 to} E0 ₇ shown in (H).

On the other hand, the model of the shape to be extracted is the direction characteristic of the normal line at M (13 in this example) evaluation points taken along the contour of the shape (represented by discretized numbers).
And the amount of movement from the previous evaluation point. The normal direction and the movement amount are called attributes of the model evaluation point, and M is called a sequence number. At this time, the model R is given by the following equation. R = {(r _m , p _m , q _m ) | m = 1 to M} 0 ≦ r _m ≦ 7 (2) In the equation (2), r _m is the direction number of the m-th evaluation point. In the example of the model shown in FIG. 18, the inside of the octagon has low density and the background has high density. Therefore, r
_m represents a normal direction having an outer direction, which is discretized into eight directions. The features obtained with this model are:
Since it is to be compared with the directional feature element obtained from the target image, it is desirable to set it at a place where the directionality is stable, that is, a linear point where the contour is less curved.

Further, in (2) (p _m, q _m)
Is the amount of movement from the (m-1) th evaluation point to the mth evaluation point. However, (p ₁ , q ₁ ) is the amount of movement from the M-th evaluation point to the first evaluation point. When the location of the M-th evaluation point is called a reference point, the cumulative amount of movement from this reference point to the m-th evaluation point is given by the formula 1.

[Equation 1] Here, the model R is possessed as knowledge or a dictionary by the processing system side prior to the processing of the input image information. As a method of creating the same, the shape or symbol to be extracted is input, and (a) the operator specifies the position of the evaluation point. (B) The processing device automatically determines (for example, at regular intervals on the contour) the evaluation point and the normal direction of the point. (C) All points of the contour are taken as evaluation points, and their normal directions are featured (in this case, p _m and q _m are within a range of ± 1). Etc. are possible. In any case, the model information is defined and stored in the processing device prior to processing the input image.

Next, storage means for holding the intermediate results of processing
The counter surface C = c (i, j) as
It Here, the number of meshes on the counter surface is the mesh on the input surface.
The number of model evaluation points m
The evaluation value for each input point is stored. here
In the expression, C represents the entire image of the counter surface c (i, j).
I, c [p_m, Q_m] For the entire image C (p_m, Q_m)
Shall mean an image that is shifted (translated) by
It One of the goals of the process according to the invention is
Directional characteristics of specified evaluation points and positional relationship between evaluation points
The shapes (symbols) that have a relation are arranged from the input image information.
It is extracted by a column operation, which will be described below. 1) Initial setting After clearing the counter surface C to 0, input the counter surface C
Overlay on surface U. Pay attention to the point (i, j) on this counter surface C
And this is the peep hole W_ijIf you call, peep hole W _ijIs input
It is placed at the position of the point (i, j) in the image information. FIG. 19
(A) shows the position of the counter surface C at the time of initial setting and the reference point.
Peep hole W located_ijIs indicated by a small circle. At this time, cow
The input surface C and the input surface U completely overlap each other.

2) For the model evaluation points m = 1 to M, the following processing is sequentially executed. 2-1) the entire moving counter surface C (p _m, q _m) shifted (parallel movement). This process is expressed by the following equation. C: = c [p _m, q _m] in ... (3) where the symbol ":" means substituting the computed value on the right side to the left side. This parallel movement processing is uniform for each point c (i, j) in the plane and is independent of each point, so it can be executed at high speed by parallel processing.
Here, the peephole W _ij in the initial setting is (i
+ Σp _m , j + Σq _m ).
That is, FIG. 19B shows the movement when m = 1,
The peep hole W _ij moves to the position of the evaluation point of m = 1 of the model shown by the original broken line.

2-2) Addition On the counter surface C obtained by the movement, the direction number
r_mDirection plane u based on_rmAnd specify this value as a counter
Add to the value of surface C. That is, C: = C + u_rm ... Processing such as (4) is performed. This addition process also applies to each point c in the plane.
Uniform processing for (i, j) and independent calculation of each point
Therefore, it can be executed at high speed by parallel processing. Up
Peep hole W by the processing of the expressions (3) and (4)_ijFocus on
Then, a new peep hole W_ijIs updated by the following formula. W_ij= W_ij+ U_rm(I + Σp_m, J + Σq_m) (5) As is clear from the equation (5) and FIG.
A counter surface C shown by a line and an input surface U shown by a broken line
Is added in the same position. That is, W _ijAdded to
The position indicated by i, j is not the current position.
The position at the time of initial setting shown in FIG.
are doing. Note that FIG. 19C shows the cow when m = 4 is calculated.
The position of the input surface C is shown. Current peep hole W_ijPlace of
The position is the position of the evaluation point of m = 4 at the time of initial setting. At this time
The position of the input surface U of is the position shown by the broken line,
Not in.

Focusing on the peep hole W _ij after the movement of the item 2-1) and the addition process of the item 2-2) have been performed M times, from the way of setting (p ₁ , q ₁ ), i + Σp _m Since ≡i, j + Σq _m ≡j, W _ij returns to the original position (i, j). That is, in the finally obtained _Wij , when the reference point (end point) of the model is placed at the position (i, j) of the input image,
The above-mentioned evaluation values at M evaluation points along the contour of the model, that is, the sum total of the designated direction planes are included. Since this process is performed in parallel for point (i, j),
The sum total of such evaluation amounts for all points (i, j) is obtained at the same time. Therefore, by finding a value, max {c (i, j)}, at which this sum is maximum in the counter surface C, try to put the point having this maximum value, that is, the shape of the model, at every position of the input image. ,
It is possible to know the point (i, j) that best matches among them and the evaluation value c (i, j) at that time. In addition, when it is considered that there are multiple shapes close to the model in the plane, those positions are extracted by extracting not only the point having the maximum value but also the point having a larger value than the surroundings, that is, the maximum point. To be done.

FIG. 20 shows the above processing three-dimensionally. That is, the uppermost surface C is the counter surface and the lower surface 8
The surface is the direction surface U obtained from the input image information. Each of the eight planes U has a characteristic of a directional component indicated by an arrow as shown. As the processing progresses, only the uppermost counter surface moves as shown in the figure. This movement is shown in FIG.
It corresponds to (a) to (c). At the time of addition at each evaluation point m, the value of the counter surface and the value of the directional surface are added on the peephole indicated by the vertical line in the figure. At this time, all the values of the directional surface 8 are added. No, 8 specified in the model
It is the value of one face within the face. 2-3) In addition to the above-described method, here, matching (matching) of symbol models is further performed, and a negative pattern evaluation point sequence is calculated by the following mathematical expression 2 for a category showing a matching degree of a certain value or more. Find NPH.

[Equation 2]

The flow is shown in FIG. In Equation 2, S ₀ is the original pattern of the corresponding symbol, R ₁ , R ₂ ... _Ri are the evaluation point sequence patterns of the negative symbols, and D ^k is k.
It represents a non-directional dilation operation of times, and a negative sign "-" is added. The contraction calculation is a logical product (AND) of the value of a certain point and the value of the surrounding (neighboring) in the binary image information.
Is a calculation for obtaining a new value at that point, and the expansion calculation is a logical sum (OR) calculation with the neighborhood. Further, when the pattern of the corresponding symbol and the pattern of the negative partner are superimposed, the reference point of the symbol and the point showing the maximum value of the pattern of the negative partner are aligned. The evaluation points of the negative pattern evaluation point sequence NPH will be referred to as negative evaluation points, and the evaluation points obtained up to the addition processing in the item 2-2) will be referred to as positive evaluation points. Then, when matching with the input image, addition of the counter surface is performed at the positive evaluation point and subtraction is performed at the negative evaluation point, and the processing flow thereof is shown in FIG. The flag shown in the figure is for indicating whether the evaluation point is positive (P) or negative (N).

Now, the embodiment of the present invention shown in FIG. 1 will be described. First, a model evaluation point series is created. A reference model image 1 in which a model to be recognized is written on a sheet or the like is input to the optical image input unit 3, and the grayscale image is stored in the input image memory 4. At this time, it is necessary for the input reference model image 1 to have a good image quality when creating the evaluation point series as shown in FIG. Although the reference model image 1 includes two models 21 and 22 here, more models may be included as reference models. Next, the model is cut out one by one by the cutout processing by the image calculation unit 6, and evaluation points are created. The image calculation unit 6 can perform various processes. For example, the edge operator as shown in FIGS. 16A to 16C is processed, and the edge image as a processing result is stored in the calculation image memory 5. The orientation surface conversion unit 7 separates the obtained edge image into a plurality of orientation surfaces. The number of planes in the direction may be any, but here it is eight.

On the other hand, in order to determine the evaluation point of the model, the image calculation unit 6 further binarizes the obtained edge image,
Evaluation points are determined from the contour points as shown in FIG. In the decision, there is a method of choosing such that the distance between evaluation points is constant,
It can be changed according to the purpose and target. The model creation unit 10 describes the model based on the obtained direction planes and evaluation points. As shown in FIG. 3, at the n-th evaluation point 32 of the model 22, for example, when looking at the data of the direction surface at that point, it can be seen that there is data in the second direction surface. Further, the movement amount from the immediately preceding evaluation point 31 is (x _n , y _n ).
Then, at the nth place where the model 22 is registered in the model storage unit 11, the designation flag is described as P (positive evaluation point) as shown in FIG. By making such a description for all evaluation points, the model 22
A series of positive evaluation points will be created.

The model creating unit 10 also creates a negative evaluation point (N). Models 21 and 22 as above
, It is understood that the “F” of the model 22 forms a part of the “E” of the model 21. That is, if the recognition target is "E", this is recognized by both the model 21 and the model 22, and the difference in the matching degree is very small. Therefore, when all the recognition is completed, it becomes a delicate matter whether it is “E” or “F”. Therefore, the model creating unit 10 searches for similar or inclusive ones from all combinations of models, finds a difference pattern between them, and determines this difference pattern as a negative evaluation point with N as a designated flag. I am trying to add it to the evaluation point series. FIG. 5 shows an example of creating negative evaluation points for the models 21 and 22. Here, 41 is detected as the difference pattern. In this way, the evaluation point series for all models is created, and the result is stored in the model storage unit 11.

Next, recognition is performed. Now, assuming that the recognition target 2 as shown in FIG. 6 is input, this is the model 2
It is similar to 2. This may be considered as a kind of error, but in such a case, the model storage unit 1
The model registered in 1 does not match even if recognition is performed. Therefore, in the present invention, the size of each model is changed by the model conversion unit 12 so as to be matched. This size conversion is performed for each evaluation point, and one method is to change the size of the vector connecting each evaluation point from a certain point on the image at the same rate as shown in FIG. The model is enlarged or reduced by a method of uniformly changing the size of the vector of each evaluation point. In this way, by executing the recognition by enlarging or reducing the evaluation points of the model, the recognition process becomes constant regardless of the size of the model, and therefore unknown patterns can be obtained by simply changing the size. It becomes possible to recognize. When the image quality of the recognition target image is good, the rough size 51 is estimated by a circumscribing rectangle as shown in FIG. 6, and only the large and small ranges 52 to 53 are recognized with respect to the estimated size. As a result, the recognition process can be significantly compressed.

When the recognition target 2 as shown in FIG. 8 is input, similarly to the above, the direction of the vector connecting the respective evaluation points from a certain point on the image ( The angle θ) can be changed at a constant rate, or the direction of the vector between the evaluation points can be changed uniformly.
The model shall be rotated. In the case of posture conversion, the movement amount of the evaluation point is converted and the data of the reference direction surface is also converted at the same time. In this way, if the recognition is performed for all the rotation angles, the unknown pattern can be recognized.
By the way, since the model has directional characteristics,
It has a direction stimulus like (a). Therefore, assuming that a similar direction histogram is extracted from the target image as shown in (b) of the figure, if this is compared with that of the model, it is rotated by the rotation angle 76 with respect to the model, or 74 It is considered that the rotation is performed by the rotation angle 77 with respect to 73. Therefore, the estimated rotation angle 76 is rotated to recognize only the range 78, the rotation angle 77 is rotated to recognize only the range 78, and the range 75 is similarly rotated and recognized. As a result, the recognition processing amount can be significantly reduced.

Next, consider a recognition target as shown in FIG. In this image, since the recognition target is covered by a rough background, the edge image is as shown in FIG. 8B, and many directional features are extracted in the background portion. Therefore, a simple model may be recognized at any position, and in such a case, it is difficult to distinguish it even if the negative evaluation points are taken into consideration. Therefore, an evaluation point having an omnidirectional feature that indicates the presence or absence of a model is extracted from the evaluation point sequence, and this is added to the evaluation point sequence having a directional feature to be recognized. Considering this point in the example of the model 22, after the positive evaluation point series 82 is formed on the contour line 81 of the model 22 as shown in FIG. 12, the image calculation unit 6 of FIG.
Is for the input image stored in the input image memory 4,
An image that distinguishes the model part from the background part by binarization is stored in the arithmetic image memory 5, and further thinning is performed to create an evaluation point 84 of the non-directional feature from the thin line 83. To do. In FIG. 12, “□” indicates the evaluation score of the directional feature, “◯” indicates the evaluation score of the non-directional feature, and 85 indicates the negative evaluation score.

In the recognition when an image to be recognized as shown in FIG. 11 (a) is input, a direction plane is first created, and the feature extraction unit 8 performs a process such as binarization as shown in FIG.
One non-directional characteristic surface U ′ is created. Then, according to the description of the model, the attribute of the evaluation point currently being evaluated refers to the direction plane P (positive evaluation point of the directional characteristic),
N (negative evaluation point of directional feature), p (positive evaluation point of non-directional feature), or n (negative evaluation point of non-directional feature) is determined by the designated flag, and the counter surface C is determined. The selection unit 9 switches the image to be added / subtracted. Then, the addition / subtraction calculation unit 13 adds in the case of positive evaluation points P and p,
In the case of negative evaluation points N and n, subtraction is executed respectively.
FIG. 14A shows an example in which recognition is performed only by the evaluation points of the directional features. As shown in the figure, "1" or "J" is erroneously recognized due to the strong directional characteristic of the background portion. On the other hand, FIG. 13B shows the case where the evaluation points of the non-directional features are taken into consideration, and the features of FIG. 13 can prevent erroneous recognition.

[0023]

According to the present invention, not only an unknown pattern in which an error, a similar deformation, or a posture deformation due to rotation has occurred can be recognized by one type of reference model, but the size and rotation angle can be estimated. This makes it possible to significantly reduce the amount of recognition processing. Further, it is possible to accurately recognize even a pattern on a complicated and rough background, and it is possible to avoid erroneous recognition due to a complicated background.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is an explanatory diagram illustrating an example of a reference model image.

FIG. 3 is an explanatory diagram for explaining a specific directional surface of a model.

FIG. 4 is an explanatory diagram illustrating an example of contents of a model storage unit.

FIG. 5 is an explanatory diagram illustrating an example of negative evaluation points of similar models.

FIG. 6 is an explanatory diagram for explaining an example of an input image having a similarity relationship with a model.

FIG. 7 is an explanatory diagram for explaining a size conversion method.

FIG. 8 is an explanatory diagram illustrating an example of an input image having a relationship of being rotated by a predetermined angle with respect to a model.

FIG. 9 is an explanatory diagram illustrating an example of a rotation method.

FIG. 10 is an explanatory diagram illustrating an example of a histogram of a model and an input image.

FIG. 11 is an explanatory diagram for explaining an example of a recognized image in a rough background and its edge image.

FIG. 12 is an explanatory diagram illustrating an example of a non-directional feature of a model.

FIG. 13 is an explanatory diagram illustrating an example of an omnidirectional feature screen.

FIG. 14 is an explanatory diagram for explaining the effect of the present invention.

FIG. 15 is an explanatory diagram for explaining a directional surface.

FIG. 16 is an explanatory diagram illustrating an example of an edge operator.

FIG. 17 is an explanatory diagram for describing an application example of an edge operator.

FIG. 18 is an explanatory diagram for explaining an evaluation point of a certain symbol.

FIG. 19 is an explanatory diagram for explaining a method of moving the counter surface.

FIG. 20 shows the relationship between the counter surface and the plurality of input direction surfaces in 3
It is explanatory drawing for dimensionally explaining.

FIG. 21 is a flowchart for explaining how to obtain a negative evaluation score.

FIG. 22 is a flowchart for explaining a recognition method using a positive evaluation point and a negative evaluation point.

[Explanation of symbols]

 1 Reference model image 2 Recognition target image 3 Input unit 4 Input image memory 5 Image memory for calculation 6 Image calculation unit 7 Directionalization unit 8 Feature extraction unit 9 Selection unit 10 Model creation unit 11 Model storage unit 12 Conversion unit 13 Addition / subtraction calculation Part 14 Maximum value detection part 15 Initial value setting part 16 Moving part 17 Control part

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Kazuhiko Yamamoto 1-4-1 Umezono, Tsukuba-shi, Ibaraki Electronic Technology Research Institute (72) Hirozo Yamada 1-4-1 Umezono, Tsukuba-shi, Ibaraki Electronic technology (72) Inventor, Taiichi Saito, 1-4, Umezono, Tsukuba-shi, Ibaraki Electronic Technology Research Institute (72) Inventor, Katsumi Hosokawa, 1 Fuji-machi, Hino-shi, Tokyo Fuji-Facom Control Co., Ltd.

Claims (2)

[Claims] 1. An image recognition apparatus for performing image recognition by comparing a directional feature extracted from an input image by parallel calculation with a model of a predefined shape, and inputting a recognition target image and a reference model image. An image input means, an image storage means for storing the input image, a model of a form to be extracted is expressed as a series of evaluation points, and a value relating to the directional feature to be referred to for each evaluation point and the evaluation A model storage unit that defines the amount of movement from the immediately preceding evaluation point in the series of points and stores the value and the amount of movement related to the defined directional characteristics in advance, and each of the models has a high degree of matching. At this time, additional means for adding to the evaluation point sequence of the model as information for distinguishing the value and the movement amount regarding the directional characteristics of the evaluation points corresponding to those difference patterns from the similar model, Direction plane creating means for creating a plurality of direction planes expressing the strength of the direction component by extracting the directional features of the input image; and counter plane storage means for storing a counter plane for each point of the input image. A moving means for moving the counter surface in parallel according to a moving amount defined for each evaluation point, and the counter surface moved by the moving means, on the counter surface, of the direction surface obtained by the value relating to the directional characteristic. Addition / subtraction means for adding or subtracting data, repeating means for repeating the parallel movement and addition / subtraction in the order of the evaluation point series of the model, and each point on the counter surface after the end of the parallel movement and addition / subtraction iterations. Detection means for detecting the shape and position of the model from the distribution of the sum of the directional surfaces for each, first conversion means for converting the evaluation point series into a similar model, Second conversion means for converting the evaluation point series into a rotated model; first estimation means for estimating the size of the target image; second estimation means for estimating the rotation direction and angle of the target image; An image recognition device comprising: 2. A model of a shape to be extracted is expressed as a series of evaluation points, a value relating to non-directionality indicating the presence or absence of a model for each evaluation point, and a movement amount from the immediately preceding evaluation point in the series. 2. The image recognition apparatus according to claim 1, further comprising: extracting means for defining the above, and adding the extracting means for recognizing the quantity in combination with the evaluation point series having the directional characteristic.